Analysis of medical images for earlier than early detection of cancer

Abstract

There is provided a method of identifying likelihood of cancer in a subject, comprising: feeding an input image of a subject depicting a radiological lesion with at least one of: (i) a diameter less than about 10 millimeters (mm), (ii) is visually similar to neighboring tissues, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from neighboring tissues surrounding and / or in proximity to the radiological lesion, and (iii) includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features, into a machine learning model trained to analyze the radiological lesion, and obtaining an indication of likelihood of cancer or non-cancer for the radiological lesion from the machine learning model.

Description

[0001] ANALYSIS OF MEDICAL IMAGES FOR EARLIER THAN EARLY

[0002] DETECTION OF CANCER

[0003] RELATED APPLICATION

[0004] This application claims the benefit of priority of U.S. Provisional Patent Application No. 63 / 471,026 filed on June 5, 2023, the contents of which are incorporated herein by reference in their entirety.

[0005] BACKGROUND

[0006] The present invention, in some embodiments thereof, relates to cancer detection and, more specifically, but not exclusively, to system and methods for analyzing images for detecting cancer.

[0007] Cancer screening using imaging involves using medical imaging technologies like mammograms, CT scans, and MRIs to detect cancer early in asymptomatic individuals. These methods help identify tumors, abnormal growths, or other cancer indicators at an early stage. Regular imaging screenings are vital for early diagnosis and improving treatment outcomes. Different imaging techniques are recommended based on the type of cancer and individual risk factors.

[0008] SUMMARY

[0009] According to a first aspect, a computer implemented method of identifying likelihood of cancer in a subject, comprises: feeding an input image of a subject depicting a radiological lesion with at least one of: (i) a diameter less than about 10 millimeters (mm), (ii) is visually similar to neighboring tissues, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from neighboring tissues surrounding and / or in proximity to the radiological lesion, and (iii) includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features; into a machine learning model trained to analyze the radiological lesion; and obtaining an indication of likelihood of cancer or non-cancer for the radiological lesion from the machine learning model.

[0010] According to a second aspect, a computer implemented method of training a machine learning model for identifying likelihood of cancer in a subject, comprises: creating a training dataset comprising a plurality of records, wherein a record comprises a sample image of a sample individual depicting a radiological lesion with diameter less than about 10 mm, and a ground truth indicating that the radiological lesion of the sample image is cancer or non-cancer, wherein the ground truth is obtained according to a label assigned to the radiological lesion in a second image of the sample individual obtained a time interval after the sample image, the radiological lesion depicted in the second image corresponds to the radiological lesion depicted in the sample image, and has a larger diameter and / or is more enhanced than the radiological lesion depicted in the sample image, and training the machine learning model on the training dataset for generating a likelihood of cancer or non-cancer in response to being fed an input image of a subject depicting a radiological lesion with diameter less than about 10 mm.

[0011] According to a third aspect, a system for identifying likelihood of cancer in a subject, comprises: at least one processor executing a code for: feeding an input image of a subject depicting a radiological lesion with at least one of: (i) a diameter less than about 10 millimeters (mm), (ii) is visually similar to neighboring tissues, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from neighboring tissues surrounding and / or in proximity to the radiological lesion, and (iii) includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features; into a machine learning model trained to analyze the radiological lesion; and obtaining an indication of likelihood of cancer or non-cancer for the radiological lesion from the machine learning model.

[0012] In a further implementation form of the first, second, and third aspects, the benign radiological characteristics include at least one of: indicating at least one morphological characteristic, and indicating at least one straining characteristic by contrast material.

[0013] In a further implementation form of the first, second, and third aspects, the radiological lesion of the input image is further at least one of: associated with a negative biopsy, statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the input image, and non-specific wherein the input image includes a plurality of other radiological lesions statistically similar to the radiological lesion.

[0014] In a further implementation form of the first, and third aspects, the machine learning model is trained on a training dataset comprising a plurality of records, wherein a record comprises a sample image of a sample individual depicting the radiological lesion with diameter less than about 10 mm, and a ground truth indicating that the radiological lesion of the sample image is cancer or non-cancer, wherein the ground truth is obtained according to a label assigned to the radiological lesion in a second image of the sample individual obtained a time interval after the sample image, the radiological lesion depicted in the second image corresponds to the radiological lesion depicted in the sample image, and has a larger diameter and / or is more enhanced than the radiological lesion depicted in the sample image.

[0015] In a further implementation form of the second aspect, the radiological lesion of the sample image is visually similar to neighboring tissues surrounding and / or in proximity to the radiological lesion, and the radiological lesion of the second image is visually distinct in comparison to neighboring tissues surrounding and / or in proximity to the radiological lesion.

[0016] In a further implementation form of the second aspect, visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from the neighboring tissues, and visually distinct comprises features extracted from the radiological lesion are statistically different than features extracted from the neighboring tissues.

[0017] In a further implementation form of the first, second or third aspect, the feature extracted from the radiological lesion and / or neighboring tissues include at least one of: contour, pixel intensity, edges, texture, shape, enhancement, kinetic, curveology and / or deep radiomics.

[0018] In a further implementation form of the second aspect, the radiological lesion of the sample image includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features.

[0019] In a further implementation form of the first, second or third aspect, the benign radiological characteristics include at least one of: indicating at least one morphological characteristic, and indicating at least one straining characteristic by contrast material.

[0020] In a further implementation form of the first, second, or third aspect, the radiological lesion of the sample image is further at least one of: associated with a negative biopsy, statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the sample image, and non-specific wherein the sample image includes a plurality of other radiological lesions statistically similar to the radiological lesion.

[0021] In a further implementation form of the first, second, and third aspects, the radiological lesion of the input image and / or the sample image is visually similar to neighboring tissues within the breast, and the radiological lesion of the second image is visually distinct in comparison to neighboring tissues within the breast.

[0022] In a further implementation form of the first, second, and third aspects, visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from the neighboring tissues, and visually distinct comprises features extracted from the radiological lesion are statistically different than features extracted from the neighboring tissues. In a further implementation form of the first, second, and third aspects, the feature extracted from the radiological lesion and / or neighboring tissues include at least one of: contour, pixel intensity, edges, texture, shape, enhancement, kinetic, curveology and / or deep radiomics.

[0023] In a further implementation form of the first, second, and third aspects, the diameter of the radiological lesion of the input image and / or the sample image is less than about 5 mm.

[0024] In a further implementation form of the first, second, and third aspects, the radiological lesion of the input image and / or the sample image comprises an enhanced foci.

[0025] In a further implementation form of the first, second, and third aspects, the sample image and the second image overlap at least in a region that includes the radiological lesion, wherein a location of the radiological lesion in a body of the subject depicted in the second image substantially matches the location of the radiological lesion in the body of the subject depicted in the sample image.

[0026] In a further implementation form of the first, second, and third aspects, the second image is captured at least 6 months after the sample image.

[0027] In a further implementation form of the first, second, and third aspects, further comprising: for the sample image, extracting a first patch including that includes mostly the radiological lesion, and a second patch larger than the first patch that includes the first patch and surrounding tissue, wherein the record includes a combination of the first patch and the second patch.

[0028] In a further implementation form of the first, second, and third aspects, the machine learning model is implemented as a neural network including a first component of a plurality of convolutional layers designed to be fed an input corresponding the first patch and generate a first set of features extracted from the first patch, a second component of a plurality of convolutional layers designed to be fed an input of the second patch and generate a second set of features extracted from the second patch, and a third component designed to be fed a combination of the first set of features and the second set of features and generate a classification indicative of cancer.

[0029] In a further implementation form of the first, second, and third aspects, the training dataset includes at least one record of a second type wherein the ground truth indicates that the radiological lesion is non-cancer, wherein the ground truth is obtained from a second image of the sample individual obtained a time interval after the sample image, the second image depicting the radiological lesion substantially changed from the radiological lesion depicted in the sample image in terms of diameter and / or enhancement.

[0030] In a further implementation form of the first, second, and third aspects, the record further comprises a combination of the sample image and tabular data of clinical and / or radiological and / or demographic features, and further comprising extracting image features from the sample image, concatenating the image features with the tabular data, and training the machine learning model on the concatenated features.

[0031] In a further implementation form of the first, second, and third aspects, the cancer comprises breast cancer associated with a risk factor, wherein the input image and / or sample image and / or second image are captured using a breast MRI protocol.

[0032] In a further implementation form of the first, second, and third aspects, the input image and / or sample image and / or second image depict at least one breast with background parenchymal enhancement of dense tissue of a woman under about 30 years old that hides the radiological lesion in the sample image and / or the input image.

[0033] In a further implementation form of the first, second, and third aspects, further comprising extracting a first patch from the input image that includes mostly the radiological lesion, and a second patch larger than the first patch that includes the first patch and surround tissue, wherein feeding the input image comprises feeding a combination of the first patch and second patch into the machine learning model.

[0034] In a further implementation form of the first, second, and third aspects, the input image comprises a first image and a second image, wherein the first patch and the second patch are extracted from each of the first image and the second image, and further comprising extracting features from the first patch and the second patch from each of the first image to obtain a plurality of vectors, concatenating the plurality of vectors to form a single vector, wherein feeding comprises feeding the single vector.

[0035] In a further implementation form of the first, second, and third aspects, the first image comprises a conventional DCE MRI image and the second image comprises a second ultrafast DCE MRI image.

[0036] In a further implementation form of the first, second, and third aspects, the first patch and the second patch of each of the first image and the second image are extracted by feeding the first image and the second image into a detector model trained to identify potential lesions.

[0037] In a further implementation form of the first, second, and third aspects, further comprising including clinical and / or demographic features in the single vector.

[0038] In a further implementation form of the first, second, and third aspects, further comprising dividing the input image into a plurality of patch pairs including a first patch and a second patch larger than the first patch and including additional environment not depicted in the first patch, wherein feeding the input image comprises iteratively feeding a patch pair of the plurality of patch pairs of the image. In a further implementation form of the first, second, and third aspects, feeding comprises feeding a combination of the input image and tabular data including at least one of: clinical, demographic and radiological features.

[0039] In a further implementation form of the first, second, and third aspects, at least one sub- DCE (dynamic contrast enhanced) image, created by subtracting at least one pre-contrast image from at least one time-point post-contrast image.

[0040] In a further implementation form of the first, second, and third aspects, at least one timepoint post-contrast image comprises the 2ndtime-point post-contrast image.

[0041] In a further implementation form of the first, second, and third aspects, the machine learning model comprises a detector, and the indication comprises a detected location of the radiological lesion on the input image, wherein for the second image, the radiological lesion is identified and a plurality of anatomical landmarks in proximity to the identified radiological lesion are identified, for the sample image, the plurality of anatomical landmarks corresponding to the second image are identified, the first image and the second image are registered based on the corresponding plurality of anatomical landmarks, the identified radiological lesion of the second image is mapped to the first image, and the machine learning model is trained for detecting a location of the radiological lesion on the input image.

[0042] In a further implementation form of the first, second, and third aspects, the diameter of the radiological lesion of the input image and / or the sample image is less than about 5 mm.

[0043] In a further implementation form of the first, second, and third aspects, the radiological lesion of the input image and / or the sample image comprises an enhanced foci.

[0044] In a further implementation form of the first, second, and third aspects, in the sample image and the second image overlap at least in a region that includes the radiological lesion, wherein a location of the radiological lesion a body of the subject depicted in the second image substantially matches the location of the radiological lesion in the body of the subject depicted in the sample image.

[0045] In a further implementation form of the first, second, and third aspects, the second image is captured at least 6 months after the sample image.

[0046] In a further implementation form of the first, second, and third aspects, further comprising: for the sample image, extracting a first patch including mostly the radiological lesion, and a second patch larger than the first patch that includes the first patch and surrounding tissue, wherein the record includes a combination of the first patch and the second patch.

[0047] In a further implementation form of the first, second, and third aspects, the machine learning model is implemented as a neural network including a first component of a plurality of convolutional layers designed to be fed an input corresponding the first patch and generate a first set of features, a second component of a plurality of convolutional layers designed to be fed an input of the second patch and generate a second set of features, and a third component designed to be fed a combination of the first set of features and the second set of features and generate a classification indicative of cancer.

[0048] In a further implementation form of the first, second, and third aspects, further comprising: creating at least one record of a second type wherein the ground truth indicates that the radiological lesion is benign, wherein the ground truth is obtained from the second image of the sample individual obtained a time interval after the sample image, the second image depicting the radiological lesion substantially unchanged from the radiological lesion depicted in the sample image in terms of diameter and / or enhancement.

[0049] In a further implementation form of the first, second, and third aspects, the record further comprises a combination of the sample image and tabular data of clinical and / or radiological and / or demographic features, and further comprising extracting image features from the sample image, concatenating the image features with the tabular data, and training the machine learning model on the concatenated features.

[0050] In a further implementation form of the first, second, and third aspects, the cancer comprises breast cancer associated with a risk factor, wherein the sample image and / or second image and / or input image is captured using a breast MRI protocol.

[0051] In a further implementation form of the first, second, and third aspects, the input image and / or sample image and / or second image depict at least one breast with background parenchymal enhancement of dense tissue of a woman under about 30 years old that hides the radiological lesion in the sample image and / or the input image.

[0052] In a further implementation form of the first, second, and third aspects, the sample image and / or second image and / or input image each comprise an MRI image including at least one sub- DCE (dynamic contrast enhanced) image, created by subtracting at least one pre-contrast image from at least one time-point post-contrast image.

[0053] In a further implementation form of the first, second, and third aspects, the at least one time-point post-contrast image comprises the 2ndtime-point post-contrast image.

[0054] In a further implementation form of the first, second, and third aspects, further comprising: for the second image, segmenting the radiological lesion and identifying a plurality of anatomical landmarks in proximity to the segmented radiological lesion, for the sample image, identifying the plurality of anatomical landmarks corresponding to the second image, registering between the first image and the second image based on the corresponding plurality of anatomical landmarks, mapping the segmented radiological lesion of the second image to the first image, and wherein the machine learning model is trained for detecting a location of the radiological lesion on the input image.

[0055] Unless otherwise defined, all technical and / or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention pertains. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the invention, exemplary methods and / or materials are described below. In case of conflict, the patent specification, including definitions, will control. In addition, the materials, methods, and examples are illustrative only and are not intended to be necessarily limiting.

[0056] BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING(S)

[0057] Some embodiments of the invention are herein described, by way of example only, with reference to the accompanying drawings. With specific reference now to the drawings in detail, it is stressed that the particulars shown are by way of example and for purposes of illustrative discussion of embodiments of the invention. In this regard, the description taken with the drawings makes apparent to those skilled in the art how embodiments of the invention may be practiced.

[0058] In the drawings:

[0059] FIG. 1 is a block diagram of components of a system for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention;

[0060] FIG. 2 is a flowchart of a method of detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention;

[0061] FIG. 3 is a flowchart of a method of training a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention;

[0062] FIG. 4 is a schematic depicting an exemplary architecture of a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention;

[0063] FIG. 5 is presents an example of consecutive MRIs of three representative BC cases in whom an abnormality could be visualized in their prior MRI, in accordance with some embodiments of the present invention; FIG. 6 includes a table of radiological features and tumor pathological characteristics of all 53 BC cases participating in the study described herein, in accordance with some embodiments of the present invention;

[0064] FIG. 7 includes a table of a comparison between two consecutive scans (diagnosis and prior scans) for cancer and cancer-free patients included in the study, in accordance with some embodiments described herein;

[0065] FIG. 8 is a confusion matrix of the performance of the machine learning model used in the study for classification of very early lesions / abnormalities in prior MRI scans to the diagnosis scans of breast cancer, in accordance with some embodiments of the present invention;

[0066] FIG. 9 is a table of radiological and pathological characteristics of successfully and unsuccessfully classified abnormalities in prior scans of cancer patients of the study, in accordance with some embodiments of the present invention;

[0067] FIG. 10 is a table of lesion characteristics of successfully and unsuccessfully classified abnormalities in prior scans of cancer patients included in the study, in accordance with some embodiments of the present invention;

[0068] FIG. 11 depicts representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study, in accordance with some embodiments of the present invention;

[0069] FIG. 12 depicts additional representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study, in accordance with some embodiments of the present invention;

[0070] FIG. 13 depicts yet additional representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study, in accordance with some embodiments of the present invention; and

[0071] FIG. 14 is a flowchart of another exemplary method for training and / or inference by a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention.

[0072] DETAILED DESCRIPTION

[0073] The present invention, in some embodiments thereof, relates to cancer detection and, more specifically, but not exclusively, to system and methods for analyzing images for detecting cancer.

[0074] As used herein, the term cancer, lesion, and tumor may be interchanged.

[0075] As used herein, a set of images is used for training a machine learning model. A first image may also be referred to herein as a sample image, an earlier image, and / or a prior image. A second image captured a time interval after the first image may also referred to herein as a second image, a later image, and / or a follow-up image.

[0076] It is to be understood that embodiments described herein with respect to breast cancer are provided as a not necessarily limiting example, as other types of cancers may be detected. For example, embodiments described herein may be adapted to detect lung cancer nodules within lung tissue, such as small nodules that may initially look benign but are cancerous. Other types of cancer that may be detected using embodiments described herein include, for example, brain cancer, and kidney cancer.

[0077] It is to be understood that embodiments described herein with respect to MRI are provided as a not necessarily limiting example, as other types of images may be detected, for example, CT, ultrasound, and x-rays. Images described herein may include, for example, 3D images, 2D slices of 3D images, time sequences (e.g., different stages of contrast), and / or individual 2D images.

[0078] It is to be understood that embodiments described herein with respect to genetic mutations (e.g., BRCA1 / BRCA2) are provided as a not necessarily limiting example, as cancer may be due to other risk factors, such as other types of genetic mutations and / or other causes which are not necessarily correlated with known genetic markers, such as exposure to carcinogens, such as tobacco, asbestos, certain chemicals, and radiation, or personal history and / or family history.

[0079] It is to be understood that embodiments described herein are not necessarily limited to high-risk subjects (e.g., BRCA carriers, and / or women who have a personal and / or family history of breast cancer), but may also be applied for detecting and / or classifying lesions, optionally subcentimeter lesions, in low and / or average risk subjects, for example, as a tool for subjects that are being screened with imaging (e.g., breast MRI screening).

[0080] It is to be understood that embodiments described herein with respect to the CNN are provided as a not necessarily limiting example, as other architectures and / or parameters of other ML models may be implemented, for example, a detector architecture may be used to localize the detector cancer.

[0081] As used herein, the term artificial intelligence (Al) (e.g., processes, approaches) and machine learning (ML) (e.g., models) are used interchangeably.

[0082] Optionally, the image(s) (e.g. MRI image(s)) fed into the ML model, and / or the image(s) used for training the ML model, include additional clinical data of a subject, which may be fed in combination into the ML model, and / or included in combination with the training data, for example, age, medical history, family history, hormone state (e.g., administration), and the like. The additional clinical data may increase accuracy of the classification by the ML model. The ML model may perform classification and / or detection. There may be two ML models, one for classification, and another for detection, or a single ML model may perform both detection and classification. In detector implementations, patches may be automatically extracted from the entire tissue (e.g., breast) and fed into the ML model to identity suspicious regions where cancer may be located. Alternatively or additionally, the location of the radiological lesion on the earlier image may be found by registering corresponding anatomical features on the later image, and mapping a detected location of the radiological lesion identified on the later image back to the earlier image according to the registration.

[0083] As used herein, the diameter of the radiological lesion which is being detected is referred to as being about 10 millimeters (mm), is used as a not necessarily limiting example. The diameter may be other values, which may vary, for example, according to type of cancer, imaging modality, imaging protocol, visual similarity to surrounding tissues, and the like. Examples of other diameters include less than about 3 millimeters (mm), or 5 mm, or 7 mm, or 10 mm, or 12 mm, or 15 mm.

[0084] As used herein, a suspicious radiological lesion may refer to tissue that requires further investigation to determine whether it is cancerous or benign. For example, some tumors may appear to grow like cancer, but be benign without spreading to other parts of the body. Such benign growths may still require surgical removal, such as to avoid damage due to size and / or pressure applied by the tumor to neighboring tissue.

[0085] As used herein, the term non-cancer may refer to benign or normal tissue.

[0086] An aspect of some embodiments of the present invention relates to systems, methods, computing devices, and / or code instructions (stored on a data storage device and executable by one or more processors) for detecting cancer or non-cancer (or benign, or a suspicious lesion) that is traditionally difficult to detect on at least one image of the subject. For example, the cancer is less than about 10 mm and / or that is visually similar to neighboring tissue and / or appears more benign than non-benign (e.g., malignant, cancer), and / or includes one or more features (i)-(vi) discussed below. Such small lesions are difficult for a radiologist to visually identify, as described herein in additional detail. The image(s) may be, for example, a breast MRI of a young woman who is genetically susceptible to breast cancer undergoing an early screening examination. It is visually difficult to detect small cancer lesions in such images of such subjects, as discussed in greater detail below. Generally, cancer is detected when larger and / or more enhanced, after further growth. It is preferable to be able to detect cancer as early as possible. The image (also referred to herein as an input image) depicting a radiological lesion is fed into a machine learning model. The radiological lesion depicted in the input image is at least one of: (i) Less than about 10 mm.

[0087] (ii) Is visually similar to neighboring tissues. Visually similar may refer to features extracted from the radiological lesion are statistically similar to features extracted from neighboring tissues surrounding and / or in proximity to the radiological lesion.

[0088] (iii) Includes benign radiological characteristics. Benign radiological characteristics may include features extracted from the radiological lesion that are more correlated to benign features than to non-benign features. The correlation may be to features of benign radiological lesions which may be defined from other images of other subjects, and / or features which are known to be associated with being benign. Examples of benign radiological characteristics include at least one of: indicating at least one morphological characteristic such as round, smooth and / or non-invasive, and / or indicating at least one straining characteristic by contrast material in a pattern indicative of benign.

[0089] Alternatively or additionally, the radiological lesion may include one or more of:

[0090] (iv) Associated with a negative biopsy. The radiological lesion may still be cancer or may have become cancer since the previous biopsy.

[0091] (v) Statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the input image. For example, unchanged in size and / or morphological features from a previous screening image performed 6 months or 12 months earlier.

[0092] (vi) Non-specific. The input image may include multiple other radiological lesions which are statistically similar to the radiological lesion, for example, visually similar, similar size, similar morphological characteristics, similar extracted visual features, where similar may be within a range indicating statistical similarity.

[0093] The machine learning (ML) model generates an indication of whether the input image depicts cancer, or non-cancer, or other categories, for example, benign. In some embodiments, the ML model indicates whether the input image is suspicious for cancer, for example, a probability or representing a range of probability indicating likelihood of cancer. The ML model is trained as described herein. The patient may be treated accordingly.

[0094] An aspect of some embodiments of the present invention relates to systems, methods, computing devices, and / or code instructions (stored on a data storage device and executable by one or more processors) for training one or more ML models for detecting cancer or non-cancer (or benign, or a suspicious lesion) that is less than about 10 mm, in at least one image of a subject. A training dataset of multiple records is created. A record includes a sample image of a sample individual depicting the radiological lesion with diameter less than about 10 mm, and a ground truth indicating that the radiological lesion of the sample image is cancer or non-cancer (or suspicious) or benign, which may be obtained, for example, via a biopsy. It is noted that the radiological lesion may be initially difficult to visualize or cannot be visualized using the sample image alone, even by a radiologist trained to detect visually distinct radiological lesions which cannot or are difficult to detect without expertise. The radiological lesion in the sample image may include one or more of (i)-(vi) discussed above and herein. The ground truth is obtained from a label assigned to the radiological lesion in a second image of the sample individual obtained a time interval after the sample image. For example, the second image is obtained at least about 3 months, or 6 months, or 12 month, or 18 months, or 24 months, after the sample image. During this time, it is assumed that a cancerous or suspicious lesion would have grown and / or is more enhanced (e.g., due to increased vasculature), making it possible to accurately detect its presence. A radiological lesion depicted in the second image is identified (e.g., since it is larger and / or more enhanced than in the sample image). The radiological lesion in the sample image corresponding to the radiological lesion depicted in the second image is identified, for example, by searching in the anatomical region of interest (RO I) of the sample image corresponding to the ROI of the second image that includes the identified radiological lesion. The radiological lesion in the second image has a larger diameter and / or is more enhanced than the radiological lesion depicted in the sample image, making it possible to detect it. The machine learning model on the training dataset for generating a likelihood of cancer or non-cancer lesion or benign in response to being fed an input image of a subject depicting a radiological lesion with diameter less than about 10 mm.

[0095] The radiological lesion of the input image and / or the sample image may be visually similar to neighboring tissues. Features extracted from the radiological lesion of the input image and / or sample image may be statistically similar to features extracted from the neighboring tissues, such as within a range and / or below threshold. Features may be automatically extracted by a feature extraction process, and / or may be based on hand-crafted features. For example, contour, texture, shape, pixel intensity, edges, enhancement, kinetic, curveology, deep radiomics, and the like. The radiological lesion may be an enhanced foci, which may have an enhancement similar to neighboring tissues making it difficult to identify, for example, for breast imaging the enhanced foci is similar to the background parenchymal enhancement (BPE) of normal breast tissue. In breasts with high BPE, a small tumor may be masked. A radiologist viewing the input image and / or sample image may be unable to accurately distinguish the radiological lesion from neighboring tissues. In another example, a machine learning model trained using standard approaches on larger lesions that are visually distinct from neighboring tissues may be unable to accurately automatically classify such images and / or accurately detect the radiological lesion. The radiological lesion of the input image and / or sample image is in contrast to the radiological lesion of the second image, which is visually distinct in comparison to neighboring tissues. Features extracted from the radiological lesion of the second image are statistically different than features extracted from the neighboring tissues, such as greater than the threshold and / or external to the range. For example, different contour, different texture, different shape, different pixel intensity, different edges, different enhancement, different kinetic features, different curveology and / or different deep radiomics, and the like. The radiological lesion may be an enhanced foci, which may have an enhancement different than neighboring tissues making it easier to identify, for example, for breast imaging the enhanced foci is higher than the background parenchymal enhancement (BPE) of normal breast tissue. A radiologist viewing the second image may be able to accurately distinguish the radiological lesion from neighboring tissues. In another example, a machine learning model trained using standard approaches on larger lesions that are visually distinct from neighboring tissues may be able to accurately automatically classify such images and / or accurately detect the radiological lesion.

[0096] At least some embodiments address the technical problem of identifying a cancerous or non-cancer radiological lesion in an input image of an individual which is difficult to accurately detect or cannot be detected on the input image, for example, due to its size being small (e.g., less than about 10 mm in diameter) and / or due to it low enhancement and / or due to its similarity to surrounding tissues. At least some embodiments improve the technical field of machine learning, by training a machine learning model for classifying an image of an individual and / or for detecting in the image, a cancerous or non-cancer radiological lesion. The radiological lesion is difficult to accurately detect or cannot be detected on the input image, for example, due to its size being small (e.g., less than about 10 mm in diameter) and / or due to it low enhancement and / or due to its similarity to surrounding tissues and / or having benign looking characteristics.

[0097] At least some embodiments address the technical problem of obtaining training data for training one or more ML models for detecting cancer or non-cancer (or benign, or a suspicious lesion) that is difficult to accurately detect or cannot be detected on the input image, for example, due to its size being small (e.g., less than about 10 mm in diameter) and / or due to it low enhancement and / or due to its similarity to surrounding tissues and / or having benign looking characteristics. Since such small radiological lesions are difficult to see, and / or difficult to diagnose, labelled images are unavailable for training such ML models using standard approaches. At least some embodiments improve the technical field of machine learning, by generating training data for training one or more ML models for detecting cancer or non-cancer (or benign, or a suspicious lesion) that is difficult to accurately detect or cannot be detected on the input image, for example, due to its size being small (e.g., less than about 10 mm in diameter) and / or due to it low enhancement and / or due to its similarity to surrounding tissues and / or having benign looking characteristics.

[0098] For the case of breast cancer, the identification and classification may be influenced by the amount of background parenchymal enhancement (BPE), which is the enhancement of normal breast tissue. For example, in breasts with high BPE, a small tumor may be masked. Similar effects may occur when searching for other radiological lesions in other parts of the body, for example for lung nodules in lung tissue, since the lung tissue naturally looks nodular.

[0099] The most common scenarios that may occur when an image with a small radiological lesion is being analyze include:

[0100] (1) Unidentified by the radiologist because the radiological lesion looks similar to normal breast tissue and / or is too small to catch.

[0101] (2) Identified by radiologist. However, since the radiological lesion is unchanged from previous exams, a recommendation for short-term or normal imaging follow-up may be provided.

[0102] (3) Identified by radiologist and recommended for biopsy, which comes back negative.

[0103] At least some embodiments provide solutions to the technical problems described herein, and / or improve the technology described herein, by training a machine learning model on sample images with small and / or undiagnosed cancer foci (i.e., radiological lesions) and / or small undiagnosed benign foci (i.e., radiological lesions) and / or BPE (which makes it difficult to see the small radiological lesions). The ground truth for each sample image is obtained from another image, different from the sample image, captured after a diagnosis of the radiological lesion has been made, for example, after the cancer tumor or benign lesion was identified and biopsied. Alternatively or additionally, in the case of BPE, the next MRI captured a time interval after the sample image, was defined as normal or the suspicious lesions were biopsied and defined as normal breast tissue. In other words, the cancer is first detected using the later image where it may be biopsy-proven, then the cancer is located in a prior image where it is smaller, and used to label the earlier image.

[0104] People (e.g., women) who have a genetic mutation (e.g., of BRCA1 or BRCA2) are at a significantly higher risk for developing cancer (e.g., breast cancer). Early detection may be important for an improved prognosis, therefore these people are offered an intensive follow-up program, including a yearly image (e.g., MRI) scan. For the case of early detection of breast cancer in genetically susceptible women, although MRI is the most sensitive imaging modality for breast cancer detection, it was found that a significant number of tumors are overlooked or misinterpreted, leading to a delayed diagnosis. At least some embodiments are designed to improve breast cancer diagnosis at early stages, based on a machine learning model(s). In the “Examples” section below, the ML model described herein is shown to classify correctly (e.g., -65% in experiments) of the tumors at an early time point. These tumors were not suspected / diagnosed by the radiologists at that time point, but only at the next MRI scan. Such ML model(s) and / or associated processing may serve as an aid to the radiologists, for improving their decision-making and achieve an 'earlier than early' diagnosis of (e.g., breast) cancer in mutation (e.g., BRCA) carriers.

[0105] At least some embodiments address the technical problem of improving early detection of cancer (e.g., breast), in particular in in genetically susceptible individuals. At least some embodiments improve the technology of machine learning models, by training a machine learning model for improved early detection of cancer (e.g., breast), in particular in genetically susceptible individuals.

[0106] Female BRCA1 / BRCA2 (=BRCA) pathogenic variants (PVs) carriers are at a substantially higher risk for developing breast cancer (BC) compared with average risk population. Detection of BC at an early stage significantly improves prognosis. To facilitate early BC detection, a surveillance scheme is offered to BRCA PV carriers from age 25-30 years that includes annual MRI based breast imaging. Indeed, adherence to the recommended scheme has been shown to be associated with earlier disease stages at BC diagnosis, more in-situ pathology, smaller tumors, and less axillary involvement. While MRI is the most sensitive modality for BC detection in BRCA PV carriers, there are a significant number of overlooked or misinterpreted radiological lesions (mostly enhancing foci), leading to a delayed BC diagnosis at a more advanced stage. At least some embodiments include a ML learning model(s) that provide a more accurate classification of enhancing foci, in MRIs of BRCA PV carriers, thus reducing false-negative interpretations. Retrospectively identified foci in prior MRIs that were either diagnosed as BC or benign / normal in subsequent MRI are segmented (e.g., manually) and served as input for the ML model (e.g., a convolutional network architecture). As discussed in the “Examples” section below, the ML model was successful in classification of 65% of the cancerous foci, most of whom were triple-negative BC. Applying this scheme (e.g., routinely) may facilitate, for example, 'earlier than early' BC diagnosis in BRCA PV carriers.

[0107] BRCA1 and BRCA2 (heretofore BRCA) germline pathogenic variant (PVs) carriers are at a significantly higher risk for breast cancer (BC) with an estimated cumulative lifetime risk of 72% and 69%, respectively [1]. In addition, they may develop aggressive cancers at young ages, with peak incidence in the 41-50 or 51-60 age-groups in BRCA1 or BRCA2 PV carriers, respectively [1,2]. Tumor volume doubling time in invasive BCs is reportedly twice as high in BRCA carriers (46 and 52 days in BRCA1 and BRCA2, respectively) compared with age adjusted BRCA wild type high-risk women [3]. Thus, within 1 year, a tumor may double its volume by 4-5 times, implying a doubling of tumor diameter, a fact that may adversely affect tumor stage at diagnosis and consequent prognosis. Therefore, early detection of BC may be important for diagnosing of curable non-metastatic disease, at a stage in which the cancer cells have not yet gained the ability to metastasize and become resistant to targeted therapies [4] .

[0108] The current screening guidelines for female BRCA PVs carriers, including bi-annual breast imaging [5], have been shown to increase the rate of early diagnosis with more in-situ pathology, smaller tumors, and less axillary involvement compared with carriers who do not adhere to the suggested surveillance scheme [5-8]. MRI is considered the most sensitive modality for BC detection in BRCA PV carriers with sensitivity rates up to 96% [8-11]. Yet, several studies have shown that in retrospective analyses, in a significant number of MRI detected BCs (50-74%) localized radiological abnormalities were present in prior MRI, which was performed approximately one year before actual BC diagnosis [10,12-15]. Pages et al.

[0016] and Seo et al.

[0017] reported that in -50% of the reviewed cases (from a total of 56 and 72 consecutive pairs of MR imaging studies, respectively) a suspicious radiological finding of -1 cm in size could be retrospectively visualized in a previously performed MRI. Notably, Gubem-Merida et al.

[0018] reported rates of 60% (24 / 40 pairs of MRIs) displaying misclassified pre BC diagnosis MRI abnormal findings, and Bilocq-Lacoste et al.

[0015] reported even higher rates - 74% (57 / 77 pairs of MRIs). The false negative pre BC diagnosed MRIs in these studies were attributed to several factors: resemblance of the cancerous lesion to physiologic enhancement, small lesion size, stability in size and location of a non-mass in a postsurgical area [16,17]. Even when small lesions (usually 5 mm in size) are detected by MRI, their classification may be challenging because their morphologic and kinetic characteristics are commonly benign

[0019] .

[0109] In recent years, artificial intelligence (Al) processes are being assessed as additional tools for BC screening and diagnosis, mainly based on mammography captured data analyses [20-27]. MRI-based Al algorithms are also being evaluated for BC diagnosis, improved specificity, treatment response assessment, and anatomic segmentations of fibroglandular tissue (FGT) allowing the quantification of background parenchymal enhancement (BPE) [28-32]. One of the major challenges of MRI breast imaging analyses is characterizing sub-centimeter lesions, which often display benign appearing features, and differentiating them from abundant types of enhancing foci, for either benign etiologies or focal background parenchymal enhancement (BPE). With that in mind, Inventors developed and clinically-tested an Al network, trained to correctly classify enhancing foci detected on consecutive breast MRI studies of BRCA PV carriers, in order to facilitate ‘earlier than early’ BC diagnosis. In a study described in the “Examples” section, Inventors investigate the ability of an ML model based on at least one embodiment described herein to classify sub-centimeter breast abnormalities that were not originally detected as suspicious (e.g., by the radiologist). As shown and / or described herein, breast abnormalities can be visualized on initial MRI that are localized to the same anatomical area where the tumor was diagnosed on subsequent MRI, performed -1 year (or other time) after the initial MRI, in -60% of the cases (which may change according to the training data). Korhonen et al.

[0013] summarized some possible main reasons for false-negative errors: (1) technical - patient motion, artifacts etc. (2) perceptual - poor lesion conspicuity, subtle appearance of lesion etc. (3) cognitive - wrong interpretation. Misinterpretation of lesions was also found to increase the false-negative errors, mainly due to the presence of multiple breast lesions, prior biopsy or surgery and stability in size [15,16]. Bilocq-Lacoste et al.

[0015] reported that cancers that were undiagnosed had no specific MRI characteristics, receptor status, or risk factors such as gene mutation, chest radiation, family history and site of previous biopsy. Most of the undiagnosed cancers (51 from 77, 66%) were overlooked due to their small size and high BPE. In the study described herein, the abnormalities in the prior scans were significantly smaller than the diagnosed tumors.

[0110] Clauser et al.

[0036] reported that the rate of small foci (<0.5 cm lesions) identified by MRI in a high-risk population was 31.3% (from a total of 166 patients). An automated approach based on CAD was reported in 2016

[0018] where 71% of prior visible lesions and 31% of prior minimally visible lesions were detected in a group of 40 high risk cases. Several studies evaluated additional tools for an improved characterization of sub-centimeter breast lesions detected by breast imaging. Gibbs et al.

[0048] showed that radiomic analysis of small breast lesions is feasible; they showed significant differences between benign and malignant lesions for 53 / 133 calculated features, with high negative (>89%) and positive (>83%) predictive values. Lo Gullo et al.

[0038] showed that the combination of radiomics and machine learning improves the differentiation between benign and malignant small breast lesions in BRCA PV carriers compared with the BLRADS classification by the radiologists.

[0111] In the study described herein, an Al-based network (i.e., ML model) accurately reclassified more than half of cancerous breast MRI abnormalities in BRCA PV carriers with a low rate of false-positives. Notably, triple-negative tumors were more frequently successfully classified as cancerous compared with other histological tumor types. This may be related, for example, to the fact that BRCA1 mutated cases represented 85.7% of the successfully classified group and to the association of triple-negative breast tumors with specific MRI features [39-41]. In a study by Moffa et al.

[0039] , the majority of the triple-negative tumors appeared as regular shaped mass enhancements (round or oval) with circumscribed margins by MRI. Also a large portion of triple-negative breast cancers presented with a rim enhancement which was shown to be a positive predictor of this subtype. Importantly, triple-negative is associated with an increased angiogenesis and with a higher recurrence rate

[0039] . Therefore, the ability to detect triple-negative breast cancer at such an early stage, while abnormalities are mostly foci with no rim-enhancement, may reduce the clinical burden of the disease.

[0112] Before explaining at least one embodiment of the invention in detail, it is to be understood that the invention is not necessarily limited in its application to the details of construction and the arrangement of the components and / or methods set forth in the following description and / or illustrated in the drawings and / or the Examples. The invention is capable of other embodiments or of being practiced or carried out in various ways.

[0113] The present invention may be a system, a method, and / or a computer program product. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention.

[0114] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0115] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0116] Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.

[0117] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0118] These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0119] The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0120] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustration, and combinations of blocks in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.

[0121] Reference is now made to FIG. 1, which is a block diagram of components of a system 100 for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention. Reference is also made to FIG. 2, which is a flowchart of a method of detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention. Reference is also made to FIG. 3, which is a flowchart of a method of training a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention. Reference is also made to FIG. 4, which is a schematic depicting an exemplary architecture of a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention. Reference is also made to FIG. 5, which is presents an example of consecutive MRIs of three representative BC cases (502A, 504A, 506A) in whom an abnormality could be visualized in their prior MRI (502B, 504B, 506B), in accordance with some embodiments of the present invention. Reference is also made to FIG. 6, which includes a table 602 of radiological features and tumor pathological characteristics of all 53 BC cases participating in the study described herein, in accordance with some embodiments of the present invention. Reference is also made to FIG. 7, which includes a table 702 of a comparison between two consecutive scans (diagnosis and prior scans) for cancer and cancer-free patients included in the study, in accordance with some embodiments described herein. Reference is also made to FIG. 8, which is a confusion matrix 802 of the performance of the machine learning model used in the study for classification of very early lesions / abnormalities in prior MRI scans to the diagnosis scans of breast cancer, in accordance with some embodiments of the present invention. Reference is also made to FIG. 9, which is a table 902 of radiological and pathological characteristics of successfully and unsuccessfully classified abnormalities in prior scans of cancer patients of the study, in accordance with some embodiments of the present invention. Reference is also made to FIG. 10, which is a table 1002 of lesion characteristics of successfully and unsuccessfully classified abnormalities in prior scans of cancer patients included in the study, in accordance with some embodiments of the present invention. Reference is also made to FIG. 11, which depicts representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study, in accordance with some embodiments of the present invention. Reference is also made to FIG. 12, which depicts additional representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study, in accordance with some embodiments of the present invention. Reference is also made to FIG. 13, which depicts yet additional representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study, in accordance with some embodiments of the present invention. Reference is also made to FIG. 14, which is a flowchart of another exemplary method for training and / or inference by a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm, in accordance with some embodiments of the present invention.

[0122] System 100 may implement the acts of the method described with reference to FIGs. 2-14, optionally by a hardware processor(s) 102 of a computing device 104 executing code instructions stored in a memory 106.

[0123] Computing device 104 may be implemented as, for example, a client terminal, a server, a virtual server, a radiology workstation, a virtual machine, a computing cloud, a mobile device, a desktop computer, a thin client, a Smartphone, a Tablet computer, a laptop computer, a wearable computer, glasses computer, and a watch computer. Computing 104 may include an advanced visualization process that sometimes is add-on to a radiology workstation.

[0124] Computing device 104 may include locally stored software that performs one or more of the acts described with reference to FIGs. 2-14, and / or may act as one or more servers (e.g., network server, web server, a computing cloud, virtual server) that provides services (e.g., one or more of the acts described with reference to FIGs. 2-14 to one or more client terminals 108 (e.g., remotely radiology workstations, remote picture archiving and communication system (PACS) server, remote electronic medical record (EMR) server) over a network 110, for example, providing software as a service (SaaS) to the client terminal(s) 108, providing an application for local download to the client terminal(s) 108, as an add-on to a web browser and / or a medical imaging viewer application, and / or providing functions using a remote access session to the client terminals 108, such as through a web browser.

[0125] Different architectures based on system 100 may be implemented.

[0126] In one example, computing device 104 provides centralized services. Computing device 104 may perform training of one or more ML models 122A, as described herein. Computing device 104 may analyze locally captured images (e.g., captured by imaging device 112) by feeding the images into ML model(s) 122A, as described herein. Alternatively, training is performed by another computing device, and inference is centrally performed by computing device 104. Locally captured images may be provided to computing device 104 for centralized analysis, for example, inference by the trained ML model(s) 122A. Images may be provided to computing device 104, for example, via an API, a local application, via a PACS server, and / or transmitted using a suitable transmission protocol. The outcome of the analysis (e.g., classification of cancer and / or detection of a location of a suspected cancer) may be provided, for example, to client terminal(s) 108 for presentation on a display and / or local storage, feeding into another process, stored in an electronic medical record (e.g., hosted by server 118), and / or stored by computing device 104. In another example, computing device 104 provides centralized training of the ML model(s) 122A, using records to create one or more training datasets 122B provided by different client terminals 108 and / or servers 118. For example, training datasets originating from different hospitals, and / or for different imaging modalities (e.g., MR, CT, x-ray, ultrasound) and / or for different types of cancer (e.g., breast, lung, prostate, kidney). Respective generated ML models 122A may be provided to the corresponding remote devices (e.g., client terminal(s) 108 and / or server(s) 118) for local use. For example, each hospital uses the ML model created from their own training dataset for evaluation of new MR images captured at the respective hospital. In another example, computing device 104 provides localized services. For example, computing device 104 includes code locally stored and / or locally executed by a radiology workstation, and / or client running a radiology image viewing program. The code may be a plugin and / or add-on to the radiology image viewing program, to provide additional features of detecting cancer in images. Computing device 104 may locally train ML model(s) 122A using training dataset(s) 122B of images captured by a local imaging device 112. In another example, computing device 104 obtains trained ML model(s) 122A from another device. Computing device 104 analysis images that are locally captured by imaging device 112, as described herein. The outcomes of the analysis may be presented on a display (e.g., user interface 126) of computing device 104, locally stored, sent to another device for storage (e.g., PACS server), and / or fed into another application (e.g., for guiding a biopsy and / or radiation therapy).

[0127] Imaging device 112 captures and provides the images, which may be included in training dataset(s) 122A and / or for inference. Imaging device 112 may be for example, a MRI machine that captures MR images, a CT machine capturing CT scans, an x-ray machine capturing x-ray images, and / or ultrasound machine capturing ultrasound images. Multiple 2D slices may be obtained from a 3D image.

[0128] Training dataset(s) 122B may be stored in a data repository 114 and / or data storage device 122, for example, a storage server, a computing cloud, virtual memory, and a hard disk. Training dataset(s) 122B are used to train the ML model(s) 122A, as described herein. It is noted that training dataset(s) 122B may be stored by a server 118, accessibly by computing device 104 over network 110.

[0129] Computing device 104 may receive images and / or records of the training dataset(s) 122B from imaging device 112 and / or data repository 114 using one or more data interfaces 120, for example, a wire connection (e.g., physical port), a wireless connection (e.g., antenna), a local bus, a port for connection of a data storage device, a network interface card, other physical interface implementations, and / or virtual interfaces (e.g., software interface, virtual private network (VPN) connection, application programming interface (API), software development kit (SDK)).

[0130] Hardware processor(s) 102 may be implemented, for example, as a central processing unit(s) (CPU), a graphics processing unit(s) (GPU), field programmable gate array(s) (FPGA), digital signal processor(s) (DSP), and application specific integrated circuit(s) (ASIC). Processor(s) 102 may include one or more processors (homogenous or heterogeneous), which may be arranged for parallel processing, as clusters and / or as one or more multi core processing units.

[0131] Memory 106 (also referred to herein as a program store, and / or data storage device) stores code instruction for execution by hardware processor(s) 102, for example, a random access memory (RAM), read-only memory (ROM), and / or a storage device, for example, non-volatile memory, magnetic media, semiconductor memory devices, hard drive, removable storage, and optical media (e.g., DVD, CD-ROM). For example, memory 106 may store image processing code 106 A that implement one or more acts and / or features of the method described with reference to FIGs. 2-14 and / or training code 106B that trains one or more of the ML models described herein.

[0132] Computing device 104 may include a data storage device 122 for storing data, for example, one or more trained ML models 122 A as described herein and / or one or more training datasets 122B as described herein. Data storage device 122 may be implemented as, for example, a memory, a local hard-drive, a removable storage device, an optical disk, a storage device, and / or as a remote server and / or computing cloud (e.g., accessed over network 110). It is noted that 122A- B may be stored in data storage device 122, with executing portions loaded into memory 106 for execution by processor(s) 102.

[0133] ML model(s) 112A described herein may be implemented using a suitable architecture designed to process images, for example, the exemplary neural network architecture described herein, and / or other architectures, such as one or more neural networks of various architectures (e.g., detector, convolutional, fully connected, deep, U-net, encoder-decoder, recurrent, graph, and combination of multiple architectures), for example, a 2D UNet, a 3D VNet, and the like..

[0134] Computing device 104 may include a network interface 124 for connecting to network 110, for example, one or more of, a network interface card, a wireless interface to connect to a wireless network, a physical interface for connecting to a cable for network connectivity, a virtual interface implemented in software, network communication software providing higher layers of network connectivity, and / or other implementations. Computing device 104 may access one or more remote servers 118 using network 110, for example, to obtain and / or provide training dataset(s) 116, an updated version of code 106A, training code 106B, and / or the trained ML model(s) 122A.

[0135] It is noted that data interface 120 and network interface 124 may exist as two independent interfaces (e.g., two network ports), as two virtual interfaces on a common physical interface (e.g., virtual networks on a common network port), and / or integrated into a single interface (e.g., network interface). Computing device 104 may communicate using network 110 (or another communication channel, such as through a direct link (e.g., cable, wireless) and / or indirect link (e.g., via an intermediary computing device such as a server, and / or via a storage device) with one or more of:

[0136] * Client terminal(s) 108, for example, when computing device 104 acts as a server providing image analysis services (e.g., SaaS) to remote radiology terminals, for analyzing remotely obtained images using the trained ML model(s) 122A. Training dataset(s) 122B may be created based on images received from one or more client terminals 108.

[0137] * Server 118, for example, implemented in association with a PACS, which may store image for training dataset(s) 122B and / or may store captured images for inference. Training dataset(s) 122B may be created from images stored by server 118.

[0138] * Imaging device 112 and / or data repository 114 that store images acquired by imaging device 112. The acquired images may be analyzed and / or fed into trained ML model(s) 122A for inference. Training dataset(s) 122B may be created based on images obtained from one or more imaging devices 112.

[0139] Computing device 104 and / or client terminal(s) 108 and / or server(s) 118 include and / or are in communication with a user interface(s) 126 that includes a mechanism designed for a user to enter data and / or view the outcome of the ML model. Exemplary user interfaces 126 include, for example, one or more of, a touchscreen, a display, a keyboard, a mouse, and voice activated software using speakers and microphone.

[0140] Referring now back to FIG. 2, at 202, one or more machine learning models are accessed and / or trained.

[0141] An exemplary approach for training the machine learning model(s) is described herein, for example, with reference to FIG. 3.

[0142] Machine learning models may be trained, for example, for processing images captured by different imaging modalities (e.g., CT or MRI), images captured by different imaging machines (e.g., different manufacturers, different models), images captured using different imaging protocols (e.g., MRI breast protocol), image of different parts of the body for detecting different cancers (e.g., breast for breast cancer, lung for lung cancer), and the like. Alternatively, the machine learning model(s) is trained to analyze more specific images, for example, captured by a certain imaging modality, captured by a certain imaging machine, a certain imaging protocol, a certain part of the body, and the like.

[0143] The machine learning model may be implemented according to the inputs and / or outputs such as defined by the training dataset, for example, a binary classifier, a multi-class classifier, a detector, one or more neural networks of various architectures (e.g., convolutional, fully connected, deep, encoder-decoder, recurrent, transformer, graph, combination of multiple architectures), support vector machines (SVM), logistic regression, k-nearest neighbor, decision trees, boosting, random forest, TabPFN, XGBoost, a regressor and the like.

[0144] In an example in which two patches are extracted from the input image (e.g., as described herein in additional detail), the machine learning model may be implemented as a neural network including a first, second, and third component. The first component includes multiple convolutional layers designed to be fed an input corresponding to the first patch. The first component generates a first set of features. The second component includes multiple convolutional layers designed to be fed an input of the second patch and generate a second set of features. The third component is designed to be fed a combination of the first set of features and the second set of features and generate a classification indicative of cancer, or non-cancer, or benign, or suspicious. An exemplary architecture is described with reference to FIG. 4.

[0145] Referring now back to FIG. 4, exemplary architecture 402 of a machine learning model for detecting cancer in an input image depicting a radiological lesion of less than about 10 mm is depicted. The machine learning model is implemented as a convolutional neural network (CNN). The model is trained to differentiate between the categories of cancer (e.g., tumor) and non-cancer (e.g., non-tumor) ROIs. It is noted that other categories may be used, and / or three or more categories may be used, for example, a suspicious lesion. Two patches 404 and 406 per abnormality or BPE regions served as input; patch 404 includes the regions immediately around the abnormality or BPE (resized to 24x24) and patch 406 includes a larger field of view (FOV) region including the abnormality’s or BPE's (128x128) surroundings. Patch 406 may capture global spatial information and smaller local details. Architecture 402 includes a first component 408 A and a second component 408B applied to respective patches 404 and 406. Each component 408A and 408B include a repeated application of a 3x3 convolution, followed by a rectified linear activation function (ReLU) and a 2x2 max pooling operation with stride 2 for down-sampling for feature extraction. The feature vectors generated by components 408A and 408B are flattened and concatenated to a single feature vector 410. The single feature vector is fed as input into a multi- layer-perceptron classifier 412. The output is pushed through a soft-max activation to provide a valid distribution, such a classification 414 into the category of cancerous or non-cancerous (or other categories).

[0146] At 204, an input image of a subject is accessed.

[0147] The input image may depict a radiological lesion with a diameter less than about 10 mm, or 5 mm, or other values as described herein. The radiological lesion may be visually similar to neighboring tissues, such that a radiologist cannot accurately detect the radiological lesion, or has entirely missed the radiological lesion, for example, the radiological lesion is “hidden” due to its similarity to neighboring tissues.

[0148] The radiological lesion in the input image may further depict and / or include one or more * Includes benign radiological characteristics. Benign radiological characteristics may include features extracted from the radiological lesion that are more correlated to benign features than to non-benign features. The correlation may be to features of benign radiological lesions which may be defined from other images of other subjects, and / or features which are known to be associated with being benign. Examples of benign radiological characteristics include at least one of: indicating at least one morphological characteristic such as round, smooth and / or non-invasive, and / or indicating at least one straining characteristic by contrast material in a pattern indicative of benign.

[0149] * Associated with a negative biopsy. The radiological lesion may still be cancer or may have become cancer since the previous biopsy.

[0150] * Statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the input image. For example, unchanged in size and / or morphological features from a previous screening image performed 6 months or 12 months earlier.

[0151] * Non-specific. The input image may include multiple other radiological lesions which are statistically similar to the radiological lesion, for example, visually similar, similar size, similar morphological characteristics, similar extracted visual features, where similar may be within a range indicating statistical similarity.

[0152] The input image may be captured as part of a screening program to find cancer in its early stages, in particular in people with genetic mutations making them more susceptible to develop cancer. For example, for the case of breast cancer screening, the input image may be an MRI captured using a breast MRI protocol to search for breast cancer. The breast(s) depicted in the image may depict a background parenchymal enhancement of dense tissue of a woman under about 30 years old, which masks the radiological lesion in the input image.

[0153] The input image may include at least one sub-DCE (dynamic contrast enhanced) MRI image, created by subtracting at least one pre-contrast image from at least one time-point postcontrast image, optionally the 2ndtime-point post-contrast image.

[0154] At 206, tabular data may be accessed. It is noted that tabular data may refer to non-image data, such as numbers, text, labels, and the like, which are not necessary in table format.

[0155] Examples of tabular data include clinical and / or radiological features, for example, patient age, genetic status, breast cancer history (family or personal), the number of days between two scans per patient, background parenchymal enhancement (BPE) scoring, and BIRADS score etc.

[0156] At 208, the input image and / or tabular data may be processed.

[0157] Optionally, the input image is divided into multiple regions, such as patch pairs. The regions may be neighboring and / or overlapping. The regions may be created by sliding a window across the image. The patch pairs may include a smaller patch and a larger patch larger than the smaller patch and including additional environment not depicted in the smaller patch. The smaller match may be sized according to the expected size of the radiological lesion. For example, for a lesion less than about 10 mm, for a square patch, the size of the patch may be, for example, about 10-20 mm, or 15-25 mm, and the like. The larger may be, for example, about 100-150 mm, or about 50-100 mm, or other values. The two patches may capture global spatial information (larger patch) and smaller local details (smaller patch).

[0158] Optionally, features are extracted from the image, for example, kinetic, curveology and / or deep radiomics and / or other features described herein. Kinetic features may be computed, for example, as described with reference to Equations (1) and (2) in the Examples section. The number of extracted features may be, for example, about 100-200, or about 100-500, or about 200-300, and the like. A principal component analysis (PCA) may be applied to reduce the dimensionality of the features extracted from the image (i.e., image features). The reduced image features obtained from the PCA are concatenated with the set of tabular data (e.g., clinical features) collected from the patient.

[0159] At 210, the image and / or tabular data is fed into the machine learning model.

[0160] Optionally, for the implementation of patch pairs, the pair of patches may be simultaneously fed into the machine learning model designed to have two inputs for accepting the two patches, for example, as described with reference to FIG. 4.

[0161] Optionally, for the implementation of dividing the input image into regions, patches extracted from each region may be iteratively fed into the machine learning model. The regions enable searching for radiological lesions across the image. Feeding patches extracted from regions may increase accuracy of detection by providing image patches where the radiological lesion takes up a larger portion of the patch. Alternatively or additionally, the input image is fed into an initial detector model which may be inaccurate. The detector model may generated multiple bounding boxes where cancer may be likely. Patch pairs may be extracted based on the bounding boxes generated by the detector model. The extracted patch pairs may be fed into the machine learning model described herein for more accurate detection.

[0162] Optionally, for the implementation of a combination of the image and tabular data, the combined features vector may be fed as input to the machine learning model, for example a classifier such as Random Forest, TabPFN, XGBoost, and the like.

[0163] At 212, an outcome is obtained from the machine learning model. The outcome may be, for example, a classification category indicating likelihood of cancer, or non-cancer. Another potential classification category is suspicious. Alternatively or additionally, the machine learning model may generate a probability value indicating likelihood of cancer, and / or accuracy of the classification category. Alternatively or additionally, the machine learning model may be a detector, generating a bounding box on the input image indicating the location where the radiological lesion is found.

[0164] The outcome may be provided, for example, presented on a display, stored in a medical record of the subject (and / or on a data storage device), stored as metadata associated with the image, forwarded to another device (e.g., administrative server), and / or forwarded to another automated process such as another machine learning model.

[0165] At 214, further action may be taken according to the outcome generated by the machine learning model. For example, an alert may be presented on a display. A radiologist may inspect the image based on the alert.

[0166] Optionally, the patient may be treated according to the outcome. For example, a biopsy of the tissue corresponding to the location of the radiological lesion is obtained, the tissue may be surgically removed, radiotherapy may be applied, chemotherapy and / or biological therapy may be administered, and / or another image may be taken at a shorter interval than otherwise indicated (e.g., in another about 1 month, or 3 months).

[0167] Referring now back to FIG. 3, at 302, a sample image of a subject is accessed. The sample image may be an earlier image, which may be captured prior to a second image which was used to make a diagnosis.

[0168] The sample image may depict a radiological lesion with a diameter less than about 10 mm, or 5 mm, or other values as described herein. The radiological lesion may be visually similar to neighboring tissues, such that a radiologist cannot accurately detect the radiological lesion, or has entirely missed the radiological lesion, for example, the radiological lesion is “hidden” due to its similarity to neighboring tissues.

[0169] The radiological lesion in the sample image may further depict and / or include one or more of:

[0170] * Includes benign radiological characteristics. Benign radiological characteristics may include features extracted from the radiological lesion that are more correlated to benign features than to non-benign features. The correlation may be to features of benign radiological lesions which may be defined from other images of other subjects, and / or features which are known to be associated with being benign. Examples of benign radiological characteristics include at least one of: indicating at least one morphological characteristic such as round, smooth and / or non-invasive, and / or indicating at least one straining characteristic by contrast material in a pattern indicative of benign. * Associated with a negative biopsy. The radiological lesion may still be cancer or may have become cancer since the previous biopsy.

[0171] * Statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the input image. For example, unchanged in size and / or morphological features from a previous screening image performed 6 months or 12 months earlier.

[0172] * Non-specific. The input image may include multiple other radiological lesions which are statistically similar to the radiological lesion, for example, visually similar, similar size, similar morphological characteristics, similar extracted visual features, where similar may be within a range indicating statistical similarity.

[0173] The sample image may be selected due to its lack of available diagnosis, and / or lack of ability to make an accurate diagnosis (e.g., above a threshold) by a human radiologist and / or classifier trained using standard approaches.

[0174] The sample image may be captured as part of a screening program to find cancer in its early stages, in particular in people with genetic mutations making them more susceptible to develop cancer. For example, for the case of breast cancer screening, the input image may be an MRI captured using a breast MRI protocol to search for breast cancer. The breast(s) depicted in the image may depict a background parenchymal enhancement of dense tissue of a woman under about 30 years old, which masks the radiological lesion in the sample image.

[0175] The sample image may include at least one sub-DCE (dynamic contrast enhanced) MRI image, created by subtracting at least one pre-contrast image from at least one time-point postcontrast image, optionally the 2ndtime-point post-contrast image.

[0176] The radiological lesion of the sample image may be visually similar to neighboring tissues. Features extracted from the radiological lesion of the sample image may be statistically similar to features extracted from the neighboring tissues, such as within a range and / or below threshold. Features may be automatically extracted by a feature extraction process, and / or may be based on hand-crafted features. For example, contour, texture, shape, pixel intensity, edges, enhancement, kinetic, curveology, deep radiomics, and the like. The radiological lesion may be an enhanced foci, which may have an enhancement similar to neighboring tissues making it difficult to identify, for example, for breast imaging the enhanced foci is similar to the background parenchymal enhancement (BPE) of normal breast tissue. In breasts with high BPE, a small tumor may be masked. A radiologist viewing the input image and / or sample image may be unable to accurately distinguish the radiological lesion from neighboring tissues. In another example, a machine learning model trained using standard approaches on larger lesions that are visually distinct from neighboring tissues may be unable to accurately automatically classify such images and / or accurately detect the radiological lesion.

[0177] At 304, a second image, also referred to herein as a diagnosis image, is captured.

[0178] The second image depicts the radiological lesion in a state which enables a diagnosis, for example, the radiological lesion is larger and / or more enhanced and / or has changed it visual appearance to be distinct from the neighboring tissue.

[0179] The second image may be captured a significant amount of time after the sample image, long enough to enable the radiological lesion to change to the state which enables the diagnosis, for example, at least about 1, 3, 6, 9, 12, 18, 24 months, or other values.

[0180] The second image may be captured using the same imaging modality, imaging device, and / or imaging protocol as the sample image.

[0181] The sample image and the second image overlap at least in one region that includes the radiological lesion. A location of the radiological lesion the body of the subject depicted in the second image substantially matches the location of the radiological lesion in the body of the subject depicted in the sample image. The locations may be registered, for example, as described herein.

[0182] The radiological lesion in the second image may be visually distinct in comparison to the radiological lesion in the sample image. Features extracted from the radiological lesion of the second image are statistically different than features extracted from the neighboring tissues, such as greater than the threshold and / or external to the range. For example, different contour, different texture, different shape, different pixel intensity, different edges, different enhancement, different kinetic features, different curveology and / or different deep radiomics, and the like. The radiological lesion may be an enhanced foci, which may have an enhancement different than neighboring tissues making it easier to identify, for example, for breast imaging the enhanced foci is higher than the background parenchymal enhancement (BPE) of normal breast tissue. A radiologist viewing the second image may be able to accurately distinguish the radiological lesion from neighboring tissues. In another example, a machine learning model trained using standard approaches on larger lesions that are visually distinct from neighboring tissues may be able to accurately automatically classify such second images and / or accurately detect the radiological lesion.

[0183] At 306, ground truth data for the sample image is obtained.

[0184] The ground truth for the sample image is obtained from the diagnosis made for the second image. The ground truth may be obtained according to a label assigned to the radiological lesion in the second image. It is noted that as described herein the diagnosis cannot be made or is difficult to make for the sample image without reference to the second image, for example, as discussed herein in additional detail. The diagnosis may be obtained, for example, automatically by accessing data (e.g., electronic medical record, histology database, biopsy database, and the like) and / or by manual input. The diagnosis using the second image may be made, for example, via a biopsy, surgical resection, histological examination, radiological example, and / or by another machine learning model which may be trained using standard approaches on images with large and / or visually distinct radiological lesions labelled with the diagnosis.

[0185] At 308, tabular data may be accessed. It is noted that tabular data may refer to non-image data, such as numbers, text, labels, and the like, which are not necessary in table format.

[0186] Examples of tabular data include clinical and / or radiological features, for example, patient age, genetic status, breast cancer history (family or personal), the number of days between the two sample and seconds per patient, background parenchymal enhancement (BPE) scoring, and BIRADS score etc.

[0187] At 310, the sample image and / or second image and / or tabular data may be processed.

[0188] Optionally, the sample image and the second image are registered. Registration may be performed by identifying anatomical landmarks on the sample image, and the corresponding anatomical landmarks on the second image, for example, by extracting features from the sample image and the second image, and registering based on matching anatomical landmarks and / or features. The anatomical landmarks on the second image may be in proximity to an identified location of the radiological lesion. The radiological lesion may be identified and / or segmented, for example, manually and / or automatically by a segmentation model and / or based on vision processing approaches. The anatomical landmarks on the sample image corresponding to the anatomical landmarks on the second image may be determined, for example, using a feature matching approach that matches features extracted from the anatomical landmarks. A mapping function (e.g., transformation matrix) that maps the second image to the sample image may be computed based on the registered anatomical landmarks and / or features.

[0189] When the location of the radiological lesion is not known on the sample image, the mapping function enables mapping the location of the radiological lesion from the second image to the sample image. The radiological lesion on the second image may be segmented, and / or a boundary box may be positioned over the radiological lesion. The segmentation and / or boundary box may be mapped to the sample image by applying the mapping function.

[0190] The locations of the radiological lesions on the sample image and / or second image may be used for training a detector implementation of the machine learning model for detecting the location of the radiological lesion on input images, even when the radiological lesion is small (e.g., less than about 10 mm or 5 mm) and / or visually similar to neighboring tissues. Optionally, the sample image and / or second image are divided into multiple regions, such as patch pairs. The regions may be neighboring and / or overlapping. The regions may be created by sliding a window across the images. The patch pairs may include a smaller patch and a larger patch larger than the smaller patch and including additional environment not depicted in the smaller patch. The smaller match may be sized according to the expected size of the radiological lesion. For example, for a lesion less than about 10 mm, for a square patch, the size of the patch may be, for example, about 10-20 mm, or 15-25 mm, and the like. The larger may be, for example, about 100-150 mm, or about 50-100 mm, or other values. The two patches may capture global spatial information (larger patch) and smaller local details (smaller patch).

[0191] Optionally, features are extracted from the sample image and / or second image, for example, curveology and / or deep radiomics and / or other features described herein. The number of extracted features may be, for example, about 100-200, or about 100-500, or about 200-300, and the like. A principal component analysis (PCA) may be applied to reduce the dimensionality of the features extracted from the images (i.e., image features). The reduced image features obtained from the PCA are concatenated with the set of tabular data (e.g., clinical features) collected from the patient.

[0192] At 312, a record is created for the patient.

[0193] The record may include the sample image, and one or more ground truths. The ground truth may include the diagnosis and / or the classification category, for example, cancer, benign, and / or suspicious. The ground truth may include the segmentation of the radiological lesion and / or boundary box encompassing the radiological lesion, which may be obtained by mapping the segmentation and / or boundary box of the radiological lesion of the second image, as described herein.

[0194] Alternatively or additionally, the record may include one or more of the following: tabular data, image features extracted from the image, concatenation of features, and / or other data described herein.

[0195] At 314, one or more features described herein with reference to 302-312 may be iterated, for creating multiple records. The multiple records may be included in one or more training datasets.

[0196] Iterations may be implemented for images of different patients.

[0197] Alternatively or additionally, iterations may be implemented based on image augmentation approaches applied to the images of the same patient, to increase the size and / or diversity of the training dataset. This may enable training the ML model to classify radiological lesions from images derived from different imaging (e.g., MRI) machines, and / or with various image qualities. Examples of image augmentation that may be applied to the sample image and / or second image include; random rotations, flips, crops, random shifts to the segmentation and / or bounding boxes of the two images, such that the two images are not necessarily centrally aligned.

[0198] Alternatively or additionally, iterations may be implemented based on pairs of features extracted from the sample image.

[0199] At 316, one or more machine learning models are trained on the training dataset(s). Exemplary architectures of the machine learning model are described herein.

[0200] Referring now back to FIG. 14, the method described with reference to FIG. 14 may be an adaptation of the method described with reference to FIG. 2 and / or FIG. 3 for processing two (or more) images of different types, such as two breast MRI protocols, including a conventional dynamic contrast enhanced (DCE) 1402A and if available an ultrafast DCE 1402B. Images 1402A and 1402B may be co-registered per patient. If an ultrafast DCE scan 1402B is not available then the process may use the conventional DCE and / or another type of image.

[0201] A real-time object detection process (e.g., detector model, image processing) may be applied to the two types of images 1402 A and 1402B separately to identify potential lesion candidates. Bounding-boxes generated by the detection process where cancer is likely to be found may be tagged 1404A-B.

[0202] Two different sized bounding boxes, also referred to herein as patches, may be created per suspicious lesion in both image types: a first bounding box that mostly surrounds the lesion 1406A- B and a second bounding box that further includes the lesion environment 1408A-B. Each suspicious lesion is depicted within four ROIs (i.e., patches) from which features may be extracted. The extracted features for each image may first be concatenated first into a respective vector for each image 1410A-B, which are then further concatenated into a single vector 1412.

[0203] The feature vector 1412 may be integrated with clinical and / or demographic features 1414 (e.g., as described herein). The full multimodal feature vector created from the integration may be used for lesion classification 1416, for example, into one of three categories: malignant, benign or normal breast tissue, optionally according to a risk scale. The result may include colored maps of the breast slices, including the percentage of risk per suspicious lesion. Various embodiments and aspects of the present disclosure as delineated hereinabove and as claimed in the claims section below find experimental and / or calculated support in the following examples.

[0204] EXAMPLES

[0205] Reference is now made to the following examples, which together with the above descriptions illustrate some embodiments of the invention in a non limiting fashion. Inventors performed a study (also referred to herein as experiments) using patient data and / or images, for evaluating one or more embodiments including the ML model(s) described herein.

[0206] In this study Inventors presented the ability to accurately classify sub-centimeter MR detected breast abnormalities that were not suspected by the radiologist and were subsequently diagnosed as cancer on average of 1 year later. Embodiments based on the approaches described herein may potentially facilitate earlier BC detection in high-risk women.

[0207] 1. Materials and Methods

[0208] 1.1. Study population

[0209] This study was approved by the institutional review board and the need to obtain a written informed consent was waived, given the study outline and retrospective nature.

[0210] A retrospective search was performed in the dataset of the Meirav high-risk Clinic repository at Sheba Medical Center, for all BRCA PV carriers with MRI detected BC and a prior MRI examination in the preceding 18 months, between 2012 and 2021. Diagnosis was based on biopsy obtained pathological report and the location of the tumor was determined by MRI. Cancer- free BRCA PV carriers seen at the same clinic during the study period with consecutive MRIs, with at least one year clinical and radiological follow-up were retrieved, and served as controls.

[0211] Breast MRI was performed at 1.5 Tesla (Signa Excite HDX, GE Healthcare) using a dedicated double breast coil equipped with eight channels. Dynamic contrast enhanced (DCE)- MRI protocol was obtained via axial vibrant multiphase 3D DCE T1 -weighted sequence, prior and four times after an automated injection of contrast agent bolus [0.1 ml / kg at 2 ml / sec Dotarem, (gadoterate meglumine, Guebet)] followed by a 20 ml saline flush. The first post-contrast images were centered at 1:25 minutes after injection and the rest were centered every 2 minutes following the first images, such that the delayed images were centered at 7:35 minutes after injection. The parameters used for the DCE-MRI were as follows: echo time (TE) = 2.6 ms, repetition time (TR) = 5.4 ms, flip angle = 15°, bandwidth = 83.3 kHz, matrix = 512 x 364, field of view (FOV) = 340 mm and slice thickness = 2 mm.

[0212] The pre-contrast images were subtracted from the 2ndtime-point post-contrast images to create sub-DCE images. These were collected from the prior MRIs of all patients.

[0213] 1.2. MR images, clinical, radiological and pathological data analysis

[0214] Patients' consecutive MRIs; 'prior' (~1 year before diagnosis) and 'diagnosis' were collected along with their clinical data, pathology of the diagnosed tumors and image radiology features including: age, time between consecutive scans, exact BRCA PV, breast imaging reporting and data system (BI-RADS) score, and BPE grade in prior scans, tumor size at diagnosis, tumor histopathological type, histological grade and immunohistochemistry results for hormonal receptor expression. Lesions at diagnosis and abnormalities in the prior scans (if existed) were retrospectively morphologically characterized (focus, mass, non-mass) according to the BI-RADS lexicon [acr.org / birads] by an experienced breast radiologist (MSL).

[0215] 1.3. Lesion segmentation and morphological / kinetic assessment

[0216] The prior MRIs were used for analysis and the diagnosis MRIs and biopsies served as ground-truth for cancer / cancer-free labeling and for identification of the tumor area and its delineation.

[0217] Manual segmentation was performed on picture archiving and communication system (PACS). In the cancer patients: the central slice of the tumor was manually marked on the sub- DCE image of the diagnosis MRI by DA in agreement with radiologist MSL. Manual coregistration between the two consecutive scans in each patient was performed in PACS and then the corresponding region of the known cancerous tumor was identified in the prior scan based on anatomical landmarks. In cases in which the region comprised of an enhancing abnormality, its central slice was manually marked. In the cancer-free women, if the prior MRIs had a reported lesion / abnormality, it was marked for control. Otherwise, a prominent enhancing focus (regarded as BPE) was chosen and marked. The corresponding regions in the following MRI scans were carefully identified and marked as well, if existed. In all cases, the segmentation was copied from the subtraction image to the raw data including the pre-contrast, the first and the fourth postcontrast time point images. Then, kinetic features were calculated according to:

[0218] Equation (1): Initial enhancement = ((SE in 1st time point)-(SE in pre contrast)) / (SE in pre contrast) *100

[0219] Equation (2): Delayed phase = ((SE in 4th time point)-(SE in 1st time point)) / (SE in 1st time point) *100, where SE denotes signal enhancement. Initial enhancement was categorized as 'slow' when smaller than 50%, 'medium' when between 50-100% and 'fast' when larger than 100%. Delayed phase was defined as 'persistent' for a larger increase than 10%, 'plateau' when between a decrease of 10% and an increase of 10%, 'washout' for a decrease of more than 10%

[0033] .

[0220] Segmented lesions and abnormalities in both prior and diagnosis MRIs were morphologically characterized into 3 categories: focus, mass and non-mass according to BI-RADS lexicon [acr.org / birads]. Lesion size was manually measured in PACS based on the largest diameter.

[0221] 1.4. Convolutional neural network (CNN) architecture

[0222] The ML model categorized between tumor and non-tumor RO Is. Two patches per abnormality or BPE regions served as input; (1) the regions immediately around the abnormality or BPE (resized to 24x24) and (2) a larger field of view (FOV) region including the abnormality’s or BPE's (128x128) surroundings. This allowed to capture both global spatial information and smaller local details

[0034] . The network architecture that was used is described with reference to FIG. 4. The training was performed using an Adam optimizer with a learning rate of le-4(1 x 10-4) that minimizes a binary cross entropy loss function. The feature extractors were pre-trained on a Cifar-19 dataset for an improved initialization of the model's weights. Cross-validation was performed with a leave-one-out procedure.

[0223] 1.5. Statistics:

[0224] The data were statistically analyzed using SPSS 29.0 (Chicago, IL, USA). T-tests or Chi- square tests were used to assess differences between independent groups. Repeated measures ANOVA with a Greenhouse-Geisser correction was used to test the change in lesion characteristics between prior and diagnosis scans in the cancerous and non-cancerous groups. Time of scan was used as the within-subject factor and group as the between- subjects factor. Paired t-tests were applied as post-hoc analysis in the cases where a significant interaction between time and group was found. Significance was set at p<0.05.

[0225] 2. Results:

[0226] Overall, the study group encompassed 53 biopsy-proven BC in BRCA PV carriers with a pre diagnosis MRI available for analysis (367.6 days between scans on average). As controls, 53 cancer-free BRCA PV carriers with two consecutive MRIs performed during the study period were analyzed.

[0227] In retrospective visualization, a radiological abnormality could be detected in previous MRIs of 32 / 53 BC patients (60.4%) in the same anatomical region where tumor was subsequently located.

[0228] Referring now back to FIG. 5, examples of consecutive MRIs of three representative BC cases (502A, 504A, 506 A) in whom an abnormality could be visualized in their prior MRI (502B, 504B, 506B) are depicted. The time difference between the two scans was 357, 418, and 387 days, respectively. Diagnosed BCs in three patients (502A, 504A, 506A) and their corresponding radiological abnormalities as appeared in the previous MRIs (502B, 504B, 506B), are shown. Tumors and abnormalities are marked by a white circle 508A-B, 510A-B, and 512A-B. The morphological and kinetic features are shown below the images in tables 514, 516, and 518. Case 1 - 74 years old BRCA1 PV carrier, diagnosed with grade 3 triple negative ductal carcinoma in situ (DCIS) (502A). Linear non mass enhancement at the prior MRI (502B) was unchanged throughout several previous MRIs (not presented), and therefore reported as BLRADS 2. Case 2 - 64 years old BRCA1 PV carrier, diagnosed with grade 3 triple negative invasive ductal carcinoma (IDC) (504A). Enhancing focus at the prior MRI was misinterpreted as intramammary lymph node and reported as BLRADS 2 (504B). Case 3 - 67 years old BRCA1 PV carrier, diagnosed with grade 3 triple negative IDC (506 A). Enhancing focus at the prior MRI was not detected by the radiologist and reported as BLRADS 2 (506B).

[0229] Referring back to FIG. 6, table 602 presents radiological features and tumor pathological characteristics of all 53 BC cases participating in the study described herein. Continuous data are mean ± SD, with ranges in parenthesis. Categorical data are number of patients with percentages in parenthesis. IDC = invasive ductal carcinoma, DCIS = ductal carcinoma in situ, HER2 = human epidermal growth factor receptor, HR = hormone receptor. Of 32 patients in whom a radiological abnormality was present in previous MRI imaging, 5 (15.6%) cases were reported as BLRADS 0 due to that abnormality or due to dense breasts, indicating that more imaging information is required (mammography and ultrasound) in order to make an informed medical decision. Of the 5 patients, 3 followed the recommendations, there the findings were either determined as post- surgical changes or were not detected, thus were not diagnosed at that time point. 8 / 32 (25%) patients were reported as BLRADS 3 due to unchanged findings, post-surgical changes, dense breasts or benign-looking lesions. The rest (59.4%) were reported as radiologically negative (BIRADS 1 or 2 on prior scan). Age at diagnosis, distribution of BRCA PV, days between the two scans, BLRADS and BPE in the prior scan were similar between the case group (those with BC) and the cancer-free group.

[0230] Invasive ductal carcinoma (IDC) tumors were more common in both groups compared with ductal carcinoma in-situ (DCIS), while a large number of them were high grade triple negative (44% and 60% of patients with abnormality and without abnormality in the prior MRI, respectively). Tumor size at diagnosis was significantly different between the groups (student’s t- test, p=0.01) while patients who had an abnormality in the early scan presented with larger tumor sizes than those who did not at the diagnosis MRI (12.8 mm vs. 7.8 mm on average, respectively). The distribution of BPE of the prior scans was different between the groups: in patients who had an abnormality on prior scan, BPE was distributed evenly between minimal-mild and moderate- marked, whereas in those who did not show an abnormality more MRIs showed a minimal-mild BPE score than moderate-marked (71.4% vs. 28.6%, respectively). Yet this difference was not statistically significant (p=0.08). Age of the individuals diagnosed with BC and age at analysis for the cancer-free group, days between the consecutive scans, BIRADS and BPE distributions were similar between these two groups. Mutated BRCA gene distribution was statistically significantly different (p=0.015): most cancer patients were BRCA1 PV carriers (73.6%) whereas in the cancer- free cohort an almost even representation was seen for both genes.

[0231] Referring now back to FIG. 7, table 702 includes a comparison between two consecutive scans (diagnosis and prior scans) for cancer and cancer-free patients included in the study. Continuous data are means, with ranges in parenthesis. Categorical data are numbers of patients with percentages in parenthesis. The overall mean difference of the lesion size over time was statistically significant (F(l,59) = 23.742, p<0.001), as was the interaction time*group (F(l,59) = 25.990, p<0.001). Additionally, the overall mean difference of the morphology characteristics over time (focus, mass, non-mass) was statistically significant (F(l,63) = 5.968, p=0.017) and the interaction as well (F(l,63) = 13.871, p<0.001). A paired t-test revealed significance between these features in the cancer group only showing smaller lesion size in the prior scans compared with the diagnosis scans (6.1 mm vs. 10.8 mm on average, p<0.0001), most of which were determined to be 'focus' (59.3%). The morphology type distribution was significantly different between the prior and diagnosis scans in the cancer group (p=0.0001) showing mostly 'mass' lesions at diagnosis (56.3%).

[0232] In the cancer-free cohort, almost 40% of the enhancements that were detected in the prior scans did not appear in the follow-up MRI. In cases where lesions were visualized in both consecutive MRIs, lesion size and morphology were similar.

[0233] Kinetic characteristics were similar between the consecutive scans in both the cancer and cancer-free groups.

[0234] Referring now back to FIG. 8, confusion matrix 802 depicts the performance of the machine learning model used in the study for classification of very early lesions / abnormalities in prior MRI scans to the diagnosis scans of breast cancer. Analysis based on the network architecture of the ML model described herein was used to achieve early detection of the abnormalities on prior MRIs. The network was based on a total of 85 MRI scans: prior MRIs of the 32 cancer patients who had an abnormality and prior scans of the 53 cancer-free patients. The network successfully classified 21 / 32 cancer cases (65.6%) as cancerous and 47 / 53 cancer-free cases (88.7%) as non- cancerous, as shown in confusion matrix 802. Referring now back to FIG. 9, table 902 presents radiological and pathological characteristics of successfully and unsuccessfully classified abnormalities in prior scans of cancer patients of the study. MR scan characteristics and tumor pathology and morphology were observed for the 21 successfully classified cancer cases compared with the 11 unsuccessful classified cases by the network. Interestingly, the molecular subtype distribution was significantly different between the groups (p=0.016). In the successfully classified group of patients, most tumors were triple negative (61.9%), while in the unsuccessfully classified patients, most tumors were HR+ / HER2- (50%). The tumor size showed a trend p-value of 0.05, while the correctly classified abnormalities in the prior scans had a smaller diameter compared with those which were not classified correctly (10.6 mm vs. 16.7 mm, respectively).

[0235] Referring now back to FIG. 10, table 1002 presents lesion characteristics of successfully and unsuccessfully classified abnormalities in prior scans of cancer patients included in the study. MR scan characteristics and tumor pathology and morphology were observed for the 21 successfully classified cancer cases compared with the 11 unsuccessful classified cases by the network. Morphology and kinetic characteristics were similar in both the successfully and unsuccessfully classified groups of patients. In both groups most abnormalities were ‘focus’ with a medium / fast initial enhancement and persistent delayed phase. Computer aided detection (CAD) was negative in most cases.

[0236] Referring now back to FIGs. 11-13, representative cases of successful and unsuccessful classification of the machine learning model in cancerous and benign lesions as part of the study are depicted.

[0237] In FIG. 11, four successful classifications of cancer tumors showing the abnormalities that appeared in the prior MRIs (1102A-D) and the detected tumors in the diagnostic MRIs (1102A1- Dl). Four cancer cases (1102A-A1, 1104B-B1, 1104C-C1, and 1104D-D1) of a successful classification of the ML model based on the prior MRIs are shown. In all cases an abnormality was retrospectively found in the prior scans but were not suspected at the time. Top row shows the prior MRIs of the patients (1102A-D), where arrows 1104A-D point to an enhancing abnormality. Bottom row (1102A1-D1) shows the 'diagnostic' MRIs of the patients, where arrows 1104A1-D1 point to the diagnosed cancerous tumors.

[0238] In 1102 A and 1102B, the abnormalities were detected at the time of the prior MRIs by the radiologist, however they were said to be post-surgical changes and given BI- RADS 3. Both were diagnosed ~1 year later by the 'diagnostic' MRI as IDCs. In 1102C and 1102D, there was no suspicious radiological finding at the time of the prior MRIs. The abnormalities were noted only retrospectively. Both were diagnosed by the 'diagnostic' MRIs as in- situ carcinomas. In FIG. 12, three examples of unsuccessful classifications of cancerous tumors are shown. All were given BI-RADS 2 in the prior MRIs and the radiological abnormalities were only retrospectively detected. Note that in 1202A and 1202C, the abnormalities are mildly conspicuous and particularly difficult to differentiate from the enhancing background. In all cases an abnormality was retrospectively found in the prior scans but were not suspected at the time by the radiologist. Top row (1202A-C) shows the prior MRIs of the patients, arrow 1204A-C point to an enhancing abnormality. Bottom row (1202A1-C1) shows the 'diagnostic' MRIs of the patients, arrows 1204A1-C1 point to the diagnosed cancerous tumors.

[0239] FIG. 13 shows three benign cases, including two benign cases of a successful classification of the ML model (1302A, 1302A1, 1302B, 1302B1) and one benign case of an unsuccessful classification of the ML model (1302C and 1302C1). In all cases an abnormality was retrospectively found in the prior scans. Top row (1302A-C) shows the prior MRIs of the patients, with arrows 1304A-C pointing to an enhancing abnormality. Bottom row (1302A1-C1) shows the follow-up MRIs of the patients, with arrows 1304A1-C1 arrows pointing to the benign tumors. Two of the cases were correctly classified by the ML model and one was not. The correctly classified lesions were radiologically detected in the prior MRIs, one was given BLRADS 0 and requested the patient's previous MRI to observe whether the lesion had already been present in the previous MRI (1302A) and the other was given BLRADS 2 because the lesion was known and seemed unchanged. The following MRIs of both cases were given BLRADS 2. The incorrectly classified case (1302C) was given BLRADS 4 due to non-mass enhancement that appeared in the prior scan. The patient was recommended to complete more imaging tests and a biopsy, however, the lesion was not detected under targeted ultrasound (US) nor a following MRI, therefore was finally determined as parenchymal enhancement. A follow-up MRI, almost 1 year later, showed sporadic enhancement areas which were defined as BPE and the MRI was given BLRADS 2.

[0240] Another study performed by Inventors is now described. The additional study is for accelerated diagnosis of BRCA1 / 2 related breast cancer using quantitative computational analysis of consecutive DCE-MRI Scans using a machine learning model. BRCA1 / 2 carriers are at a high- risk for developing breast cancer (BC), likely to be high-grade, aggressive have an early-age onset. They undergo an intensive surveillance, including annual MRI. This facilitates timely diagnosis, but not to the desired extent. A preliminary study performed by Inventors revealed that in -60% cases (out of 42 patients), early signs of malignancy, were already present in the earlier scan, -1 year prior to the diagnosis. Quantitative analysis of information-rich DCE-MRI with Al techniques (e.g., ML model) may identify these early clues, accelerating the diagnosis. A method for accelerated diagnosis of BRCA-related BC and / or other cancers from other causes is described herein.

[0241] Methods: The retrospective cohort included 164 breast DCE-MRIs (82 subjects, 2 consecutive scans per subject). Group 1: 42 cancer patients, 2 scans: the diagnostic and the preceding scan (-1 year before). Group2: 40 pairs of consecutive cancer-free scans. Tumors were manually delineated and used as regions of interest (ROIs). Control ROIs were generated from Group2. A set of features was extracted inside the ROIs, using the earlier scans. The features capture both spatial and temporal behavior: “deep radiomics” and “curveology” features, respectively (-250 features). Feature selection was performed using predictive and information- theoretic criteria (Area Under ROC Curve, Jansen-Shannon divergence). Performance was assessed in a series of cross-validation experiments, each training and testing multiple SVM models.

[0242] Results: A 10-feature set & RBF-SVM, successfully identified BC in 34 / 42 (-80%) cases, based on the earlier scans. The detection rate is not 100%, since in some scenarios malignancy is really absent at the earlier exam. In Group2 33% carriers were false positive.

[0243] Conclusions: The presented method uses quantitative features and ML models to flag radiologic ally suspicious regions, even before they become visually detectable.

[0244] Yet another study performed by Inventors is now described. The additional study is for ‘earlier than early’ breast cancer detection in BRCA carriers using computerized quantitative analysis of consecutive DCE-MRI scans based on a machine learning model. Early diagnosis of breast cancer (BC) in £> / ?C4 -carriers based on computational analysis of breast DCE-MRI scans, and / or other cancers from other causes, is described. BRCA1 / 2 mutation carriers are at a high-risk for developing BC, likely to be high-grade with an aggressive progression and an early age onset. A surveillance scheme including annual MRI (from 25 years of age) reduces advanced-stage diagnosis. In a retrospective visualization of breast DCE-MRI scans from -1 year prior to the diagnosis of BC in 42 BRCA-carriers, enhancing lesions were visualized in -60% of the cases. However, cancer was revealed only -1 year later at the next annual MRI scan. Inventors hypothesized that quantitative computational analysis of DCE-MRI with Al techniques (e.g., ML model) may enable focusing and flagging radiologically suspicious regions -1 year before diagnosis.

[0245] Methods: Breast DCE-MRI scans of the following BRCA-carrier groups at two time-points were considered: (1) 42 patients at BC diagnosis and -1 year before (2) 40 cancer-free patients at consecutive time-points, -1 year apart. Tumors were delineated and -250 “curveology” and “deep radiomics” features were extracted from corresponding regions in the scans from -1 year before. A selection of 10 features was performed using feature’s AUCs (Area Under ROC Curve), followed by Jansen-Shannon divergence analysis. The predictive power was tested by an SVM model on the full- and reduced-sets of features.

[0246] Results: Using the 10-feature set, in 34 / 42 (-80%) cases, BC was successfully identified in the scans -1 year prior to diagnosis. In the comparison group 13 / 40 (33%) cancer-free carriers were misclassified as positive. These preliminary results support the feasibility of “Earlier than Early” BC diagnosis in £> / ?C4 -carriers, potentially improving significantly the detection abilities of an expert’s visual inspection.

[0247] The descriptions of the various embodiments of the present invention have been presented for purposes of illustration, but are not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement over technologies found in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

[0248] It is expected that during the life of a patent maturing from this application many relevant machine learning models will be developed and the scope of the term machine learning model is intended to include all such new technologies a priori.

[0249] As used herein the term “about” refers to ± 10 %.

[0250] The terms "comprises", "comprising", "includes", "including", “having” and their conjugates mean "including but not limited to". This term encompasses the terms "consisting of" and "consisting essentially of".

[0251] The phrase "consisting essentially of" means that the composition or method may include additional ingredients and / or steps, but only if the additional ingredients and / or steps do not materially alter the basic and novel characteristics of the claimed composition or method.

[0252] As used herein, the singular form "a", "an" and "the" include plural references unless the context clearly dictates otherwise. For example, the term "a compound" or "at least one compound" may include a plurality of compounds, including mixtures thereof.

[0253] The word “exemplary” is used herein to mean “serving as an example, instance or illustration”. Any embodiment described as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments and / or to exclude the incorporation of features from other embodiments. The word “optionally” is used herein to mean “is provided in some embodiments and not provided in other embodiments”. Any particular embodiment of the invention may include a plurality of “optional” features unless such features conflict.

[0254] Throughout this application, various embodiments of this invention may be presented in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the invention. Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well as individual numbers within that range, for example, 1, 2, 3, 4, 5, and 6. This applies regardless of the breadth of the range.

[0255] Whenever a numerical range is indicated herein, it is meant to include any cited numeral (fractional or integral) within the indicated range. The phrases “ranging / ranges between” a first indicate number and a second indicate number and “ranging / ranges from” a first indicate number “to” a second indicate number are used herein interchangeably and are meant to include the first and second indicated numbers and all the fractional and integral numerals therebetween.

[0256] It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination or as suitable in any other described embodiment of the invention. Certain features described in the context of various embodiments are not to be considered essential features of those embodiments, unless the embodiment is inoperative without those elements.

[0257] Although the invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and broad scope of the appended claims.

[0258] It is the intent of the applicant(s) that all publications, patents and patent applications referred to in this specification are to be incorporated in their entirety by reference into the specification, as if each individual publication, patent or patent application was specifically and individually noted when referenced that it is to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present invention. To the extent that section headings are used, they should not be construed as necessarily limiting. In addition, any priority document(s) of this application is / are hereby incorporated herein by reference in its / their entirety.

[0259] References

[0260] 1. Kuchenbaecker, K.B.; Hopper, J.L.; Barnes, D.R.; Phillips, K.A.; Mooij, T.M.; Roos-Blom, M.J.; Jervis, S.; Van Leeuwen, F.E.; Milne, R.L.; Andrieu, N.; et al. Risks of Breast, Ovarian, and Contralateral Breast Cancer for BRCA1 and BRCA2 Mutation Carriers. JAMA 2017, 377, 2402-2416.

[0261] 2. You, C.; Xiao, Q.; Zhu, X.; Sun, Y.; Di, G.; Liu, G.; Hou, Y.; Chen, C.; Wu, J.; Shao, Z.; et al. The Clinicopathological and MRI Features of Patients with BRCA1 / 2 Mutations in Familial Breast Cancer. Gland Surg. 2021, 10, 262-272.

[0262] 3. Tilanus-linthorst, M.M.A.; Obdeijn, I.; Hop, W.C.J.; Causer, P.A.; Leach, M.O.; Pointon, L.; Hill, K.; Klijn, J.G.M.; Warren, R.M.L.; Gilbert, F.J. BRCA1 Mutation and Young Age Predict Fast Breast Cancer Growth in the Dutch , United Kingdom, and Canadian Magnetic Resonance Imaging Screening Trials. Clin Cancer Res. 2007, 13, 7357-7362.

[0263] 4. Greaves, M.; Maley, C.C. Clonal Evolution in Cancer. Nature 2012, 481, 306-313.

[0264] 5. Elezaby, M.; Lees, B.; Maturen, K.E.; Barroilhet, L.; Wisinski, K.B.; Schrager, S.; Wilke, L.G.; Sadowski, E. BRCA Mutation Carriers: Breast and Ovarian Cancer Screening Guidelines and Imaging Considerations. Radiology 2019, 291, 554-569.

[0265] 6. Bernstein-molho, R.; Kaufman, B.; Ben, M.A.; Sklair-levy, M.; Madoursky, D.; Zippel, D.; Laitman, Y.; Friedman, E. Breast Cancer Surveillance for BRCA1 / 2 Mutation Carriers e Is “ Early Detection ” Early Enough ? The Breast 2020, 49, 81-86.

[0266] 7. Guindalini, R.S.; Zheng, Y.; Abe, H.; Whitaker, K.; Toshio, F.; Walsh, T.; Schacht, D.; Kulkami, K.; Sheth, D.; Verp, M.S.; et al. Intensive Surveillance with Biannual Dynamic Contrast-Enhanced Magnetic Resonance Imaging Downstages Breast Cancer in BRCA1 Mutation Carriers. Clin. Cancer Res. 2020, 25, 1786-1794.

[0267] 8. Shraga, S.; Grinshpun, A.; Zick, A.; Kadouri, L.; Cohen, Y.; Maimon, O.; Adler- Levy, Y.; Zeltzer, G.; Granit, A.; Maly, B.; et al. “High-Risk Breast Cancer Screening in BRCA 1 / 2 Carriers Leads to Early Detection and Improved Survival After a Breast Cancer Diagnosis.” Front. Oncol. 2021, 77, 1-10.

[0268] 9. Kriege, M.; Brekelmans, C.T.M.; Boetes, C.; Besnard, P.; Zonderland, H.M.; Obdeijn, I.; Manoliu, R.; Kok, T.; Peterse, H.L.; Tilanus-linthorst, M.M.; et al. Efficacy of MRI and mammography for breast-cancer screening in women with a familial or genetic predisposition. N Engl J Med. 2004, 351, 427-437.

[0269] 10. Lo, G.; Scaranelo, A.M.; Aboras, H.; Ghai, S.; Kulkami, S.; Fleming, R.; Bukhanov, K.; Crystal, P. Evaluation of the Utility of Screening Mammography for High-Risk Women Undergoing Screening Breast MR Imaging. Radiology 2017, 285, 36-43. 11. Warner, E.; Hill, K.; Causer, P.; Plewes, D.; Jong, R.; Yaffe, M.; Foulkes, W.D.; Ghadirian, P.; Lynch, H.; Couch, F.; et al. Prospective Study of Breast Cancer Incidence in Women With a BRCA1 or BRCA2 Mutation Under Surveillance With and Without Magnetic Resonance Imaging. J Clin Oncol. 2011, 29, 0-5.

[0270] 12. Maxwell, A. J.; Lim, Y.Y.; Hurley, E.; Evans, D.G.; Howell, A.; Gadde, S. False- Negative MRI Breast Screening in High-Risk Women. Clin. Radiol. 2017, 72, 207-216.

[0271] 13. Korhonen, K.E.; Samantha, P.; Susan, P.; Tobey, J.; Birnbaum, J. A.; Mcdonald, E.S. Breast MRI : False-Negative Results and Missed Opportunities. RadioGraphics 2021, 41, 10- 15.

[0272] 14. Gao, Y.; Reig, B.; Heacock, L.; Bennett, D.L.; Heller, S.L.; Moy, L.; Louis, S. Magnetic Resonance Imaging in Screening of Breast Cancer. Radiol Clin North Am. 2021, 59, 85- 98.

[0273] 15. Bilocq-Lacoste, J.; Ferre, R.; Kuling, G.; Martel, A.L.; Tyrrell, P.N.; Li, S.; Wang, G.; Curpen, B. Missed Breast Cancers on MRI in High-Risk Patients: A Retrospective Case- Control Study. Tomography 2022, 8, 329-340.

[0274] 16. Pages, EB.; Millet, I.; Hoa, D.; Doyon, F.C. Undiagnosed Breast Cancer at MR imaging: analysis of causes. Radiology 2012, 264, 40-50.

[0275] 17. Seo, M.; Cho, N.; Bae, M.S.; Koo, H.R.; Kim, W.H.; Lee, S.H.; Chu, A. Features of Undiagnosed Breast Cancers at Screening Breast MR Imaging and Potential Utility of Computer-Aided Evaluation. Korean J Radiol. 2016, 17, 59-68.

[0276] 18. Gubern-merida, A.; Vreemann, S.; Marti, R.; Melendez, J.; Lardenoije, S.; Mann, R.M.; Karssemeijer, N.; Platel, B. Automated Detection of Breast Cancer in False-Negative Screening MRI Studies from Women at Increased Risk. Eur J Radiol. 2016, 85, 472-479.

[0277] 19. Meissnitzer, M.; Dershaw, D.D.; Feigin, K.; Bernard-davila, B.; Barra, F.; Morris, E.A. MRI Appearance of Invasive Subcentimetre Breast Carcinoma : Benign Characteristics Are Common. Br J Radiol. 2017, 90, 20170102. o

[0278] 20. Lang, K.; Dustier, M.; Dahlblom, V.; Akesson, A.; Andersson, I.; Zackrisson, S. Identifying Normal Mammograms in a Large Screening Population Using Artificial Intelligence. Eur. Radiol. 2021, 31, 1687-1692.

[0279] 21. Zhou, L.; Wu, X.; Huang, S.; Wu, G.; Ye, H.; Radiology, Q.W. Detection of Breast Cancer with Mammography: Effect of an Artificial Intelligence Support System. Radiology 2020, 294, 19-28.

[0280] 22. Rodriguez-Ruiz, A.; Lang, K.; Gubem-Merida, A.; Breeders, M.; Gennaro, G.;

[0281] Clauser, P.; Helbich, T.H.; Chevalier, M.; Tan, T.; Mertelmeier, T.; et al. Stand-Alone Artificial Intelligence for Breast Cancer Detection in Mammography: Comparison With 101 Radiologists. J. Natl. Cancer Inst. 2019, 111, 916-922.

[0282] 23. Dembrower, K.; Wahlin, E.; Liu, Y.; Salim, M.; Smith, K.; Lindholm, P.; Eklund, M.; Strand, F. Effect of Artificial Intelligence-Based Triaging of Breast Cancer Screening Mammograms on Cancer Detection and Radiologist Workload: A Retrospective Simulation Study. Lancet Digit. Heal. 2020, 2, e468-e474.

[0283] 24. McKinney, S.M.; Sieniek, M.; Godbole, V.; Godwin, J.; Antropova, N.; Ashrafian, H.; Back, T.; Chesus, M.; Corrado, G.C.; Darzi, A.; et al. International Evaluation of an Al System for Breast Cancer Screening. Nature 2020, 577, 89-94.

[0284] 25. Lotter, W .; Diab, A.R.; Haslam, B.; Kim, J.G.; Grisot, G.; Wu, E.; Wu, K.; Onieva, J.O.; Boyer, Y.; Boxerman, J.L.; et al. Robust Breast Cancer Detection in Mammography and Digital Breast Tomosynthesis Using an Annotation-Efficient Deep Learning Approach. Nat. Med. 2021, 27, 244-249.

[0285] 26. Raya-Povedano, J.L.; Romero-Martin, S.; Elias-Cabot, E.; Gubem-Merida, A.; Rodriguez-Ruiz, A.; Alvarez-Benito, M. AI-Based Strategies to Reduce Workload in Breast Cancer Screening with Mammography and Tomosynthesis: A Retrospective Evaluation. Radiology 2021, 300, 57-65.

[0286] 27. Larsen, M.; Aglen, C.F.; Lee, C.I.; Hoff, S.R.; Lund-Hanssen, H.; Lang, K.; Nygard, J.F.; Ursin, G.; Hofvind, S. Artificial Intelligence Evaluation of 122969 Mammography Examinations from a Population-Based Screening Program. Radiology 2022, 303, 502-511.

[0287] 28. Xu, X.; Fu, L.; Chen, Y.; Larsson, R.; Zhang, D.; Suo, S.; Hua, J.; Member, Z. Breast Region Segmentation Using Convolutional Neural Network in Dynamic Contrast Enhanced MRI. 40thAnnual International Conference of the EMBC 2018, 750-753.

[0288] 29. Ha, R.; Chang, P.; Merna, E.; Mutasa, S.; Karcich, J.; Wynn, R.T.; Liu, M.Z. Fully Automated Convolutional Neural Network Method for Quantification of Breast MRI Fibroglandular Tissue and Background Parenchymal Enhancement. J Digit Imaging 2019, 141— 147.

[0289] 30. Comes, M.C.; Fanizzi, A.; Bove, S.; Didonna, V.; Diotaiuti, S.; Forgia, D. La; Latorre, A.; Martinelli, E.; Mencattini, A.; Nardone, A.; et al. Early Prediction of Neoadjuvant Chemotherapy Response by Exploiting a Transfer Learning Approach on Breast DCE - MRIs. Sci. Rep. 2021, 1-12.

[0290] 31. Verburg, E.; Gils, C.H. Van; Velden, B.H.M. Van Der; Bakker, M.F. Deep Learning for Automated Triaging of 4581 Breast MRI Examinations from the DENSE Trial. Radiology 2022, 29-36. 32. Abe, H.; Giger, M.L. Use of Clinical MRI Maximum Intensity Projections for Improved Breast Lesion Classification with Deep Convolutional Neural Networks. J Med Imaging 2023, 5, 014503.

[0291] 33. Kim, H.; Ko, E.Y.; Kim, K.E.; Kim, M.K.; Choi, J.S.; Ko, E.S.; Han, B.K. Assessment of Enhancement Kinetics Improves the Specificity of Abbreviated Breast MRI: Performance in an Enriched Cohort. Diagnostics 2023, 13, 136.

[0292] 34. Frid-Adar, M.; Diamant, I.; Klang, E.; Amitai, M.; Goldberger, J.; Greenspan, H. Modeling the Intra-Class Variability for Liver Lesion Detection Using a Multi-Class Patch-Based CNN. 2017, arXiv:1707.06053v2.

[0293] 35. Alahmer, H.; Ahmed, A. Hierarchical Classification of Liver Tumor from CT Images Based on Difference-of-Features (DOF). International Conference of Signal and Engineering, 2016.

[0294] 36. Clauser, P.; Cassano, E.; Nicolo, A. De; Rotili, A.; Bonanni, B.; Bazzocchi, M.; Zuiani, C. Foci on Breast Magnetic Resonance Imaging in High - Risk Women: Cancer or Not? Radiol. Med. 2016, 121, 611-617.

[0295] 37. Gibbs, P.; Onishi, N.; Sadinski, M.; Gallagher, K.M.; Hughes, M.; Martinez, D.F.; Morris, E.A.; Sutton, E.J. Characterization of Sub-1 Cm Breast Lesions Using Radiomics Analysis. J. Magn. Reson. Imaging 2019, 50, 1468-1477.

[0296] 38. Gullo, R. Lo; Daimiel, L; Saccarelli, C.R.; Bitencourt, A.; Gibbs, P.; Fox, M.J.; Thakur, S.B.; Martinez, D.F.; Jochelson, M.S.; Morris, E.A.; et al. Improved Characterization of Sub-Centimeter Enhancing Breast Masses on MRI with Radiomics and Machine Learning in BRCA Mutation Carriers. Eur Radiol 2020, 6721-6731.

[0297] 39. Mo, G.; Galati, F.; Collalunga, E.; Rizzo, V.; Amati, G.D.; Pediconi, F.; Kripa, E. Can MRI Biomarkers Predict Triple-Negative Breast Cancer? Diagnostics 2020, 15, 1090.

[0298] 40. Chen, I.E.; Lee-Felker, S. Triple-Negative Breast Cancer: Multimodality Appearance. Curr. Radiol. Rep. 2022, 11, 53-59.

[0299] 41. Li, J.; Han, X. Research and Progress in Magnetic Resonance Imaging of TripleNegative Breast Cancer. Magn. Reson. Imaging 2014, 32, 392-396.

Claims

WHAT IS CLAIMED IS:

1. A computer implemented method of identifying likelihood of cancer in a subject, comprising: feeding an input image of a subject depicting a radiological lesion with at least one of:(i) a diameter less than about 10 millimeters (mm),(ii) is visually similar to neighboring tissues, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from neighboring tissues surrounding and / or in proximity to the radiological lesion, and(iii) includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features; into a machine learning model trained to analyze the radiological lesion; and obtaining an indication of likelihood of cancer or non-cancer for the radiological lesion from the machine learning model.

2. The computer implemented method of claim 1, wherein the benign radiological characteristics include at least one of: indicating at least one morphological characteristic, and indicating at least one straining characteristic by contrast material.

3. The computer implemented method of claim 1, wherein the radiological lesion of the input image is further at least one of: associated with a negative biopsy, statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the input image, and non-specific wherein the input image includes a plurality of other radiological lesions statistically similar to the radiological lesion.

4. The computer implemented method of claim 1, wherein the machine learning model is trained on a training dataset comprising a plurality of records, wherein a record comprises a sample image of a sample individual depicting the radiological lesion with diameter less than about 10 mm, and a ground truth indicating that the radiological lesion of the sample image is cancer or non-cancer, wherein the ground truth is obtained according to a label assigned to the radiological lesion in a second image of the sample individual obtained a time interval after the sample image, the radiological lesion depicted in the second image corresponds to the radiologicallesion depicted in the sample image, and has a larger diameter and / or is more enhanced than the radiological lesion depicted in the sample image.

5. The computer implemented method of claim 4, wherein the radiological lesion of the input image and / or the sample image is visually similar to neighboring tissues within the breast, and the radiological lesion of the second image is visually distinct in comparison to neighboring tissues within the breast.

6. The computer implemented method of claim 5, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from the neighboring tissues, and visually distinct comprises features extracted from the radiological lesion are statistically different than features extracted from the neighboring tissues.

7. The computer implemented method of claim 6, wherein the feature extracted from the radiological lesion and / or neighboring tissues include at least one of: contour, pixel intensity, edges, texture, shape, enhancement, kinetic, curveology and / or deep radiomics.

8. The computer implemented method of claim 4, wherein the diameter of the radiological lesion of the input image and / or the sample image is less than about 5 mm.

9. The computer implemented method of claim 4, wherein the radiological lesion of the input image and / or the sample image comprises an enhanced foci.

10. The computer implemented method of claim 4, wherein the sample image and the second image overlap at least in a region that includes the radiological lesion, wherein a location of the radiological lesion in a body of the subject depicted in the second image substantially matches the location of the radiological lesion in the body of the subject depicted in the sample image.

11. The computer implemented method of claim 4, wherein the second image is captured at least 6 months after the sample image.

12. The computer implemented method of claim 4, further comprising: for the sample image, extracting a first patch including that includes mostly the radiological lesion, and a secondpatch larger than the first patch that includes the first patch and surrounding tissue, wherein the record includes a combination of the first patch and the second patch.

13. The computer implemented method of claim 12, wherein the machine learning model is implemented as a neural network including a first component of a plurality of convolutional layers designed to be fed an input corresponding the first patch and generate a first set of features extracted from the first patch, a second component of a plurality of convolutional layers designed to be fed an input of the second patch and generate a second set of features extracted from the second patch, and a third component designed to be fed a combination of the first set of features and the second set of features and generate a classification indicative of cancer.

14. The computer implemented method of claim 4, wherein the training dataset includes at least one record of a second type wherein the ground truth indicates that the radiological lesion is non-cancer, wherein the ground truth is obtained from a second image of the sample individual obtained a time interval after the sample image, the second image depicting the radiological lesion substantially changed from the radiological lesion depicted in the sample image in terms of diameter and / or enhancement.

15. The computer implemented method of claim 4, wherein the record further comprises a combination of the sample image and tabular data of clinical and / or radiological and / or demographic features, and further comprising extracting image features from the sample image, concatenating the image features with the tabular data, and training the machine learning model on the concatenated features.

16. The computer implemented method of claim 4, wherein the cancer comprises breast cancer associated with a risk factor, wherein the input image and / or sample image and / or second image are captured using a breast MRI protocol.

17. The computer implemented method of claim 16, wherein the input image and / or sample image and / or second image depict at least one breast with background parenchymal enhancement of dense tissue of a woman under about 30 years old that hides the radiological lesion in the sample image and / or the input image.

18. The computer implemented method of claim 1 , further comprising extracting a first patch from the input image that includes mostly the radiological lesion, and a second patch larger than the first patch that includes the first patch and surround tissue, wherein feeding the input image comprises feeding a combination of the first patch and second patch into the machine learning model.

19. The computer implemented method of claim 18, wherein the input image comprises a first image and a second image, wherein the first patch and the second patch are extracted from each of the first image and the second image, and further comprising extracting features from the first patch and the second patch from each of the first image to obtain a plurality of vectors, concatenating the plurality of vectors to form a single vector, wherein feeding comprises feeding the single vector.

20. The computer implemented method of claim 19, wherein the first image comprises a conventional DCE MRI image and the second image comprises a second ultrafast DCE MRI image.

21. The computer implemented method of claim 19, wherein the first patch and the second patch of each of the first image and the second image are extracted by feeding the first image and the second image into a detector model trained to identify potential lesions.

22. The computer implemented method of claim 19, further comprising including clinical and / or demographic features in the single vector.

23. The computer implemented method of claim 1, further comprising dividing the input image into a plurality of patch pairs including a first patch and a second patch larger than the first patch and including additional environment not depicted in the first patch, wherein feeding the input image comprises iteratively feeding a patch pair of the plurality of patch pairs of the image.

24. The computer implemented method of claim 1, wherein feeding comprises feeding a combination of the input image and tabular data including at least one of: clinical, demographic and radiological features.

25. The computer implemented method of claim 4, wherein the input image and / or sample image and / or second image each comprise an MRI image including at least one sub-DCE (dynamic contrast enhanced) image, created by subtracting at least one pre-contrast image from at least one time-point post-contrast image.

26. The computer implemented method of claim 25, wherein at least one time-point post-contrast image comprises the 2ndtime-point post-contrast image.

27. The computer implemented method of claim 4, wherein the machine learning model comprises a detector, and the indication comprises a detected location of the radiological lesion on the input image, wherein for the second image, the radiological lesion is identified and a plurality of anatomical landmarks in proximity to the identified radiological lesion are identified, for the sample image, the plurality of anatomical landmarks corresponding to the second image are identified, the first image and the second image are registered based on the corresponding plurality of anatomical landmarks, the identified radiological lesion of the second image is mapped to the first image, and the machine learning model is trained for detecting a location of the radiological lesion on the input image.

28. A computer implemented method of training a machine learning model for identifying likelihood of cancer in a subject, comprising: creating a training dataset comprising a plurality of records, wherein a record comprises a sample image of a sample individual depicting a radiological lesion with diameter less than about 10 mm, and a ground truth indicating that the radiological lesion of the sample image is cancer or non-cancer, wherein the ground truth is obtained according to a label assigned to the radiological lesion in a second image of the sample individual obtained a time interval after the sample image, the radiological lesion depicted in the second image corresponds to the radiological lesion depicted in the sample image, and has a larger diameter and / or is more enhanced than the radiological lesion depicted in the sample image; andtraining the machine learning model on the training dataset for generating a likelihood of cancer or non-cancer in response to being fed an input image of a subject depicting a radiological lesion with diameter less than about 10 mm.

29. The computer implemented method of claim 28, wherein the radiological lesion of the sample image is visually similar to neighboring tissues surrounding and / or in proximity to the radiological lesion, and the radiological lesion of the second image is visually distinct in comparison to neighboring tissues surrounding and / or in proximity to the radiological lesion.

30. The computer implemented method of claim 29, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from the neighboring tissues, and visually distinct comprises features extracted from the radiological lesion are statistically different than features extracted from the neighboring tissues.

31. The computer implemented method of claim 30, wherein the feature extracted from the radiological lesion and / or neighboring tissues include at least one of: contour, pixel intensity, edges, texture, shape, enhancement, kinetic, curveology and / or deep radiomics.

32. The computer implemented method of claim 28, wherein the radiological lesion of the sample image includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features.

33. The computer implemented method of claim 32, wherein the benign radiological characteristics include at least one of: indicating at least one morphological characteristic, and indicating at least one straining characteristic by contrast material.

34. The computer implemented method of claim 28, wherein the radiological lesion of the sample image is further at least one of: associated with a negative biopsy, statistically similar to a corresponding radiological lesion depicted in a preceding image captured a time interval prior to the sample image, and non-specific wherein the sample image includes a plurality of other radiological lesions statistically similar to the radiological lesion.

35. The computer implemented method of claim 28, wherein the diameter of the radiological lesion of the input image and / or the sample image is less than about 5 mm.

36. The computer implemented method of claim 28, wherein the radiological lesion of the input image and / or the sample image comprises an enhanced foci.

37. The computer implemented method of claim 28, wherein the sample image and the second image overlap at least in a region that includes the radiological lesion, wherein a location of the radiological lesion a body of the subject depicted in the second image substantially matches the location of the radiological lesion in the body of the subject depicted in the sample image.

38. The computer implemented method of claim 28, wherein the second image is captured at least 6 months after the sample image.

39. The computer implemented method of claim 28, further comprising: for the sample image, extracting a first patch including mostly the radiological lesion, and a second patch larger than the first patch that includes the first patch and surrounding tissue, wherein the record includes a combination of the first patch and the second patch.

40. The computer implemented method of claim 39, wherein the machine learning model is implemented as a neural network including a first component of a plurality of convolutional layers designed to be fed an input corresponding the first patch and generate a first set of features, a second component of a plurality of convolutional layers designed to be fed an input of the second patch and generate a second set of features, and a third component designed to be fed a combination of the first set of features and the second set of features and generate a classification indicative of cancer.

41. The computer implemented method of claim 28, further comprising: creating at least one record of a second type wherein the ground truth indicates that the radiological lesion is benign, wherein the ground truth is obtained from the second image of the sample individual obtained a time interval after the sample image, the second image depicting the radiological lesion substantially unchanged from the radiological lesion depicted in the sample image in terms of diameter and / or enhancement.

42. The computer implemented method of claim 28, wherein the record further comprises a combination of the sample image and tabular data of clinical and / or radiological and / or demographic features, and further comprising extracting image features from the sample image, concatenating the image features with the tabular data, and training the machine learning model on the concatenated features.

43. The computer implemented method of claim 28, wherein the cancer comprises breast cancer associated with a risk factor, wherein the sample image and / or second image and / or input image is captured using a breast MRI protocol.

44. The computer implemented method of claim 43, wherein the input image and / or sample image and / or second image depict at least one breast with background parenchymal enhancement of dense tissue of a woman under about 30 years old that hides the radiological lesion in the sample image and / or the input image.

45. The computer implemented method of claim 28, wherein the sample image and / or second image and / or input image each comprise an MRI image including at least one sub-DCE (dynamic contrast enhanced) image, created by subtracting at least one pre-contrast image from at least one time-point post-contrast image.

46. The computer implemented method of claim 45, wherein the at least one time-point post-contrast image comprises the 2ndtime-point post-contrast image.

47. The computer implemented method of claim 28, further comprising: for the second image, segmenting the radiological lesion and identifying a plurality of anatomical landmarks in proximity to the segmented radiological lesion; for the sample image, identifying the plurality of anatomical landmarks corresponding to the second image; registering between the first image and the second image based on the corresponding plurality of anatomical landmarks; mapping the segmented radiological lesion of the second image to the first image; and wherein the machine learning model is trained for detecting a location of the radiological lesion on the input image.

48. A system for identifying likelihood of cancer in a subject, comprising: at least one processor executing a code for: feeding an input image of a subject depicting a radiological lesion with at least one of:(i) a diameter less than about 10 millimeters (mm),(ii) is visually similar to neighboring tissues, wherein visually similar comprises features extracted from the radiological lesion are statistically similar to features extracted from neighboring tissues surrounding and / or in proximity to the radiological lesion, and(iii) includes benign radiological characteristics, wherein benign radiological characteristics include features extracted from the radiological lesion are more correlated to benign features than to non-benign features; into a machine learning model trained to analyze the radiological lesion; and obtaining an indication of likelihood of cancer or non-cancer for the radiological lesion from the machine learning model.

Images