Systems and methods for processing medical images

Abstract

Mammography can provide multi-view information about the health status of a person's breast, and described herein are deep learning based techniques for obtaining mammogram images associated with multiple views, extracting cross-view features from the images, and automatically determining the presence or absence of a medical abnormality based on the extracted features. Cross-view features can be determined using an encoding module of an artificial neural network (ANN), and the encoded features can be decoded using a decoding module of the ANN to generate a prediction (e.g., a classification label, a boundary shape, a segmentation mask, a probability map, etc.) about the medical abnormality.

Description

Technical Field

[0001] The field of medical image processing and analysis. Background Technology

[0002] Breast cancer accounts for a large proportion of new cancer cases and causes hundreds of thousands of deaths each year. Early screening and detection are crucial for improving breast cancer treatment outcomes and can be accomplished through mammography (mammography). While the mammographic images taken during a mammogram can provide multi-view information about the health of the breast, conventional image processing techniques may not be able to extract and fuse the information contained in the multiple views, thus preventing mammography from reaching its full potential. Summary of the Invention

[0003] This document describes systems, methods, and apparatuses associated with processing mammographic images, including digital breast tomography (DBT) images and full-field digital mammography (FFDM) images. An apparatus capable of performing such processing tasks may include at least one processor configured to acquire one or more first medical images of a person and one or more second images of the person, wherein the one or more first images may be associated with a first mammographic view of the person, and the one or more second medical images may be associated with a second mammographic view of the person. The at least one processor may also be configured to determine a first set of features associated with the one or more first medical images and a second set of features from the one or more second medical images, and to derive a third set of features based on the first and second set of features, wherein the third set of features may represent mammographic features of the person indicated by the first and second mammographic views. Based at least on the third set of features, the at least one processor may be configured to predict the presence or absence of a medical abnormality, for example, by generating an output image, which may include indications of the medical abnormality (e.g., boundary shape, segmentation mask, heatmap, etc.) or classification labels that may indicate the presence or absence of the medical abnormality.

[0004] In the example, at least one processor described herein can be configured to generate a first output image including a first indication of a medical abnormality and a second output image including a second indication of a medical abnormality, wherein the first output image may correspond to a first mammographic view and the second output image may correspond to a second mammographic view. In the example, the third feature set described herein can be determined using an artificial neural network (ANN), which may include a first encoding module and / or a second encoding module. The first encoding module may be configured to determine a first feature set associated with one or more first medical images or a second feature set associated with one or more second medical images, while the second encoding module may be configured to determine the third feature set. In the example, the ANN may also include a decoding module configured to predict the presence or absence of a medical abnormality based at least on the third feature set. Attached Figure Description

[0005] The examples disclosed herein can be understood in more detail from the following description, which is given by way of example in conjunction with the accompanying drawings.

[0006] Figure 1A and Figure 1B This is a simplified diagram illustrating examples of mammography according to some embodiments described herein.

[0007] Figure 2 This is a simplified diagram illustrating examples of machine learning (ML) frameworks for processing mammographic images such as DBT images or FFDM images, according to some embodiments described herein.

[0008] Figure 3 This is a simplified diagram illustrating an example architecture of an ML framework for processing mammographic images according to some embodiments described herein.

[0009] Figure 4 This is a flowchart illustrating an example method for processing mammographic images according to some embodiments described herein.

[0010] Figure 5 This is a flowchart illustrating an example method for training a neural network to perform one or more tasks as described with respect to some embodiments provided herein.

[0011] Figure 6 This is a simplified block diagram illustrating an example system or device for performing one or more tasks as described with respect to some embodiments provided herein. Detailed Implementation

[0012] The present disclosure is illustrated by way of example and not limitation in the accompanying drawings. A detailed description of illustrative embodiments will now be described with reference to the various drawings. Although detailed examples of possible implementations are provided in this specification, it should be noted that these details are intended to be exemplary and in no way intended to limit the scope of this application.

[0013] Mammography (mammography) is used to capture images of the breast from different views, such as cephalothorax (CC) and / or medial-lateral oblique (MLO) views. A standard mammogram may include images from multiple mammographic views, such as left CC (LCC), left MLO (LMLO), right CC (RCC), and right MLO (RMLO). Figure 1A and Figure 1B An example of mammography technique is shown, in which Figure 1A An example of full-field digital mammography (FFDM) is shown, while Figure 1B An example of digital breast tomography (DBT) is shown. Figure 1A For example, FFDM can be considered a 2D imaging modality and may involve passing a pulse train of X-rays 102 through a compressed breast 104 at an angle (e.g., perpendicular to the breast), capturing X-rays 102 on the opposite side (e.g., using a solid-state detector), and generating a 2D image 106 of the breast based on the captured signal (e.g., the captured X-rays 102 can be converted into electronic signals, which can then be used to generate the 2D image 106). Using FFDM, information about the entire breast can be incorporated into a single 2D image (e.g., 2D image 106), along with normal breast tissue (e.g., composed of...). Figure 1A (circles in the image represent) and potential lesions (e.g., caused by) Figure 1A An asterisk (representing an area on the mammogram) indicates an increased chance of overlap in the resulting mammogram image. If this overlap occurs, it can obscure lesions and increase the chance of false positives or false negatives.

[0014] Figure 1B The illustrated DBT technique can achieve or resemble the quality of a 3D imaging modality (e.g., DBT can be considered a pseudo-3D imaging modality). As shown, the technique may involve passing a pulse train of X-rays 102 through a compressed breast 104 at different angles (e.g., 0°, +15°, -15°, etc.) during a scan, acquiring a set of X-ray images of the breast at each of these angles, and reconstructing the X-ray images acquired at different angles into a series of slice images 108 (e.g., one slice image per angle), which can then be displayed individually or as a film (e.g., in a dynamic movie mode).

[0015] Figure 1A The FFDM technology shown Figure 1B Both of the DBT techniques shown can be applied to acquire medical images of the breast, which can represent different views of the breast (e.g., RCC, LCC, LMLO, and RMLO views as described herein). The information provided by these different views can then be used to differentiate normal breast tissue (e.g., by…). Figure 1A and Figure 1B The circles in the diagram represent potential lesions (e.g., caused by...). Figure 1A and Figure 1B (The asterisk in the text indicates that it enables improved diagnostic and screening results. It should be noted that, although...) Figure 1B Only three angles from which X-ray images of breast 104 were taken are shown; however, those skilled in the art will understand that more angles can be used and more images can be taken during the DBT procedure. For example, 15 images of the breast can be taken along an arc from the top and sides, and the images can be reconstructed into multiple non-overlapping slice images. Those skilled in the art will also understand that, although... Figure 1A and Figure 1B As not shown in the diagram, FFDM or DBT scans can generate different views of the two breasts, including, for example, LCC, LMLO, RCC, and RMLO views as described herein.

[0016] While the mammography techniques described herein can provide a wealth of information about potential breast conditions such as breast cancer, the amount of data generated during a mammogram (e.g., 40 to 80 slice images per view per breast during a DBT procedure) can be substantial and difficult to analyze based solely on human effort (e.g., manually by a radiologist). Therefore, in embodiments of this disclosure, artificial intelligence (AI) or deep learning (DL) based techniques are employed to automatically profil, analyze, and / or present mammogram data (e.g., FFDM and / or DBT data) in a manner that utilizes information contained in multiple views of the breast to gain a more comprehensive and accurate insight into breast health. It should be noted that the terms “artificial intelligence model,” “deep learning model,” “machine learning (ML) model,” and “artificial neural network” are used interchangeably herein, and the terms “mammogram image” or “mammography image” can include FFDM images or DDT images.

[0017] Figure 2 An example of an ML frame 200 for processing medical images of the breast (e.g., mammogram images) is illustrated. As shown, the medical image can be associated with multiple mammographic views of a person (e.g., first view 202a and second view 202b). Although in Figure 2Two mammographic views are shown, but those skilled in the art will understand that the techniques disclosed herein can also be used to process a single mammographic view (e.g., corresponding to a single input) or more than two mammographic views (e.g., corresponding to multiple inputs). As described herein, each mammographic view 202a or 202b may correspond to a corresponding set of images (e.g., one or more first images corresponding to the first view 202a and one or more second images corresponding to the second view 202b). The ML framework 200 may be configured to determine (e.g., extract) features of the images associated with the first view 202a and the second view 202b (e.g., from the respective individual images associated with view 202a or 202b), fuse and / or encode the extracted features (e.g., fuse and / or encode into one or more single-view feature representations and / or one or more cross-view feature representations), and generate an output indicating the presence or absence of a medical abnormality (e.g., lesion) in the breast, for example by decoding the encoded feature representations. For example, image features associated with the first view 202a and the second view 202b can be determined at 204a and 204b, respectively, and these specific view features (e.g., single-view features) can be used at 206 to determine a third set of features (e.g., cross-view features) that can represent mammographic features of a person as indicated by the first view 202a and the second view 202b. The cross-view features can then be decoded at 208 to generate the aforementioned output. In the example, such output may include classification labels that indicate (e.g., via Boolean values ​​or on a scale of 0 to 1) whether a medical abnormality, such as a lesion, is identified in the mammogram. In the example, such output may include one or more images having corresponding boundary shapes (e.g., bounding boxes) drawn around the medical abnormality to indicate its location. In the example, such output may include one or more segmentation masks (e.g., binary images) or one or more probability maps (e.g., heatmaps including probability values), where pixels with a first value (or within a first range) may indicate that these pixels belong to an abnormal region, and pixels with a second value (or within a second range) may indicate that these pixels do not belong to an abnormal region.

[0018] In the example, a convolutional neural network (CNN) can be trained to extract features from an input mammogram image. Such a CNN may include multiple layers, such as one or more convolutional layers, one or more pooling layers, and / or one or more fully connected layers. Each convolutional layer may include multiple convolutional kernels or filters configured to extract features from an input image (e.g., an FFDM or DBT image) received through an input channel. Following the convolution operation may be batch normalization and / or linear (or non-linear) activation, and the features extracted by the convolutional layers may be downsampled by pooling layers and / or fully connected layers to reduce feature redundancy and / or size, thereby obtaining a representation of the downsampled features (e.g., in the form of a feature vector or feature map).

[0019] In the example, an artificial neural network (ANN) (e.g., including or separate from the CNN described above) can be trained to encode and / or decode the cross-view features described herein. For example, such an ANN may include a first encoding module, a second encoding module, and / or a decoding module. The first encoding module may be trained to determine the specific view features described herein (e.g., based on local features of the input image), while the second encoding module may be trained to generate the cross-view feature representation described herein (e.g., including global features) (e.g., specific view features based on local features extracted from the input image by the CNN and / or fused from these local features). On the other hand, the decoding module may be trained to decode the features determined by the first encoding module (e.g., specific view features) and / or the features determined by the second encoding module (e.g., cross-view features), and generate the output described herein based on the decoded features.

[0020] In the example, the decoding module may include one or more up-pooling layers and one or more transposed convolutional layers, which can be configured to upsample and deconvolve the feature representation generated by the encoding module. As a result of the upsampling and deconvolution, a dense feature representation of the input image (e.g., a dense feature map) can be derived, and predictions about the presence or absence of medical abnormalities in the breast can be made based on this dense feature representation. In the example, the decoding module may be trained to acquire object query capabilities and rely on such capabilities to identify features determined by the encoding module that may be associated with medical abnormalities (e.g., these may be stored in a feature memory or feature database).

[0021] In the example, the ANN described herein can employ a transformer neural network architecture, in which one or more self-attention modules can be used to determine the features of the described specific view (e.g., a single view), one or more cross-attention modules can be used to determine the cross-view features described herein, and one or more decoding modules can be used to decode the encoded features. The self-attention and / or cross-attention modules (e.g., individual attention modules) can include a stack of layers (e.g., six identical layers), and each layer can include a set of sublayers (e.g., a multi-head self-attention mechanism and a position-wise fully connected feedforward network). Residual connections can be employed around the individual sublayers, followed by layer normalization. Thus, the output of each sublayer can be LayerNorm(x + Sublayer(x)), where Sublayer(x) can be a function implemented by the sublayer itself. The decoding module can also include a stack of layers (e.g., six identical layers) and a set of sublayers (e.g., two sublayers as described above and a third sublayer configured to perform multi-head attention on the output of the encoder stack). Residual connections can be employed around the individual sublayers, followed by layer normalization. Sublayers in the decoding module stack can be configured to prevent positional attention to subsequent positions. This masking, combined with the offset output embedding at a position, ensures that the prediction at position i can depend on (e.g., only on) known outputs at positions less than i. The attention mechanism described herein maps queries and key-value pair sets to outputs (e.g., queries, keys, values, and outputs can be vectors), and the outputs can be computed as a weighted sum of values, where the weights assigned to each value can be computed by a compatibility function of the query and its corresponding key. The fully connected feedforward network described herein can be applied individually and / or equally to each position (e.g., by performing two linear transformations with ReLU activation in between).

[0022] Using the network architecture and techniques described herein, information contained in views 202a and 202b can be fused together (e.g., a transformer can model relationships between different regions of the breast or between different views), and insights gained from one view (e.g., view 202a) can be supplemented by insights from another view (e.g., view 202b), enabling the prediction of medical abnormalities with improved accuracy. It should be noted that while the operations of 204a / 204b, 206, and 208 can be... Figure 2 The functions are shown as separate from each other (e.g., using separate neural networks to accomplish this), but those skilled in the art will understand that the functions can also be accomplished using a single neural network, such as having a CNN backbone for extracting local features from the input image and corresponding subnetworks or modules for implementing other functions (e.g., encoding and / or decoding modules).

[0023] Figure 3 An example of an ML frame 300 for processing mammogram images (e.g., DBT and / or FFDM images) is shown. Figure 2 An example of the ML framework (200). Similar to... Figure 2 Even in Figure 3 Two views (e.g., 302a and 302b) are illustrated herein, and those skilled in the art will understand that the ML framework 300 may be capable of processing only one view (e.g., as a single input) or more than two views (e.g., as multiple inputs). As described herein, each view 302a or 302b may be associated with one or more images, and the ML framework 300 may include a CNN 304 for extracting features (e.g., local features) from the image associated with view 302a or 302b and representing the extracted features with a corresponding feature map or feature vector (e.g., one feature map per image). The ML framework 300 may also include a first encoding module 306 (e.g., a self-attention or view-specific encoder) trained to fuse the features extracted by the CNN 304 into corresponding view-specific features representing mammographic features of each view. The view-specific (e.g., single-view) feature representations obtained via the first encoding module 306 may be stored (e.g., in a corresponding storage area) and provided to other modules or components of the ML framework to facilitate the prediction of medical abnormalities (e.g., lesions) in the breast.

[0024] The ML framework 300 may additionally include a second encoding module 308 (e.g., a cross-attention or cross-view encoder) trained to determine cross-view features of the breast based on specific view features fused by the first encoding module 306 and / or features extracted by the CNN 304. These cross-view features may be stored in a cross-view feature memory, from which the feature decoder 310 of the ML framework 300 may query (e.g., identify) features associated with the medical abnormality and generate an output indicating the presence or absence of the medical abnormality. As described herein, such output may include a classification label indicating whether a medical abnormality was identified in the input image. The output may also include an image (e.g., a front view of the breast) with boundary shapes (e.g., bounding boxes) drawn around the medical abnormality, or multiple images (e.g., multiple views representing the breast) with corresponding boundary shapes drawn around the medical abnormality. The images generated by the decoder 310 may also include a segmentation mask of the medical abnormality (e.g., a binary image), a heatmap associated with the medical abnormality (e.g., probability values ​​with color encoding), etc.

[0025] As explained herein, views 302a or 302b may include LCC views, RCC views, LMLO views, RMLO views, etc. The ML framework 300 is operable to generate a single output (e.g., a master view of a breast with boundary shapes drawn around a medical abnormality) based on a single input (e.g., one or more images associated with a single view) or multiple inputs (e.g., multiple views). The ML framework 300 is also operable to generate multiple outputs (e.g., multiple views) based on multiple inputs (e.g., multiple views) (e.g., multiple views of a breast with corresponding boundary shapes drawn around a medical abnormality). Similarly, as explained herein, feature encoder 306 can be trained to fuse features associated with a specific view, while feature encoder 308 can be trained to fuse features associated with multiple views. Thus, the operation of these encoders can have the effect of extracting features from the input images in a global manner. By integrating and learning from these global features, the ML framework 300 may be able to make more accurate and robust predictions than predictions made by processing individual mammographic views in isolation.

[0026] In some examples, feature decoder 310 may generate predictions based solely on cross-view features determined by feature encoder 308. In other examples, feature decoder 310 may generate predictions based on cross-view features and specific view features determined by feature encoder 306. In either case, cross-view features can be obtained by utilizing information contained in specific view features (e.g., even if feature decoder 310 generates predictions using only cross-view features). In examples, ML learning framework 300 may take images associated with multiple views as input, but may make predictions only for a subset of views. In examples, ML framework 300 may take images associated with a specific view as input and make predictions for that view (e.g., in a single-input single-output scenario). In these examples, feature encoder 308 may be unnecessary, and global features of the breast may be determined by feature decoder 310 based on specific view features provided by feature encoder 306.

[0027] Figure 4An example method 400 for processing mammographic images (e.g., DBT or FFDM images) is illustrated. Method 400 can be implemented by an ML-based system or device as described herein and may include obtaining images associated with multiple views of a human breast at 402. Method 400 may also include extracting (e.g., using a CNN backbone) corresponding features (e.g., local features) from the images at 404, and determining specific view features (e.g., global features of the individual views) and / or cross-view features of the breast based on the features extracted at 404 at 406. The method may further include decoding the cross-view features and / or specific view features of the breast at 408 and making a prediction about the presence or absence of medical abnormalities in the breast based on this decoding.

[0028] Figure 5 A flowchart is shown, illustrating the process of training a neural network (e.g., including...). Figure 2 ML framework 200 or Figure 3 The training process 500, which uses a neural network within an ML framework 300 to perform one or more tasks described herein, can be illustrated as follows: The training process 500 may include, at 502, initializing the execution parameters of the neural network (e.g., weights associated with the individual layers of the neural network), for example, by sampling from a probability distribution or by replicating the parameters of another neural network with a similar structure. The training process 500 may also include, at 504, processing one or more input images (e.g., one or more mammogram images) using the currently assigned parameters of the neural network, and at 506 making a prediction about the presence of medical abnormalities (e.g., breast disease such as lesions) in the input images. At 508, the prediction can be compared to a gold standard indicating the true state of the medical abnormality (e.g., whether breast disease is actually present) to determine a loss associated with the prediction based on a loss function. The loss function used for training can be selected based on the specific task the neural network is trained to perform. For example, if the task involves classification or segmentation of the input images, a loss function based on calculating the mean squared error between the prediction and the gold standard or based on determining the L1 or L2 norm, etc., can be used. If the task involves detecting the location of outliers around anomaly regions (e.g., by drawing bounding boxes), a loss function based on generalized intersection-union (GIOU) can be used.

[0029] At step 510, the loss calculated using one or more of the techniques described above can be used to determine whether one or more training termination criteria are met. For example, if the loss is below a threshold or if the change in loss between two training iterations is below a threshold, it can be determined that the training termination criteria are met. If the termination criteria are determined to be met at step 510, training can end; otherwise, at step 512, for example, before training returns to step 506, the currently assigned network parameters can be adjusted by backpropagating the gradient descent of the loss function through the network.

[0030] For the sake of simplicity, the training steps are depicted and described in a specific order herein. However, it should be understood that training operations can occur in various orders, simultaneously, and / or with other operations not presented or described herein. Furthermore, it should be noted that not all operations that may be included in the training method are depicted and described herein, and not all exemplified operations need to be performed.

[0031] The systems, methods, and / or apparatuses described herein may be implemented using one or more processors, one or more storage devices, and / or other suitable auxiliary devices (such as display devices, communication devices, input / output devices, etc.). Figure 6 An example device 600 is illustrated that can be configured to perform the tasks described herein. As shown, device 600 may include a processor (e.g., one or more processors) 602, which may be a central processing unit (CPU), graphics processing unit (GPU), microcontroller, reduced instruction set computer (RISC) processor, application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), physical processing unit (PPU), digital signal processor (DSP), field-programmable gate array (FPGA), or any other circuitry or processor capable of performing the functions described herein. Device 600 may also include communication circuitry 604, memory 606, mass storage device 608, input device 610, and / or communication link 612 (e.g., communication bus) through which one or more components shown in the figures exchange information.

[0032] Communication circuitry 604 can be configured to send and receive information using one or more communication protocols (e.g., TCP / IP) and one or more communication networks, including local area networks (LANs), wide area networks (WANs), the Internet, and wireless data networks (e.g., Wi-Fi, 3G, 4G / LTE, or 5G networks). Memory 606 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions that, when executed, cause processor 602 to perform one or more functions described herein. Examples of machine-readable media may include volatile or non-volatile memory, including but not limited to semiconductor memory (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)), flash memory, etc.). Mass storage device 608 may include one or more disks, such as one or more internal hard disks, one or more removable disks, one or more magneto-optical disks, one or more CD-ROMs or DVD-ROMs, etc., on which instructions and / or data may be stored for operation of processor 602. Input device 610 may include a keyboard, mouse, voice-controlled input device, touch-sensitive input device (e.g., touch screen), etc., for receiving user input from device 600.

[0033] It should be noted that device 600 can operate as a standalone device or can be connected to other computing devices (e.g., networked or clustered) to perform the tasks described herein. And even in Figure 6 Only one example of each component is shown in the figure, and those skilled in the art will understand that device 600 may include multiple instances of one or more components shown in the figure.

[0034] Although this disclosure has been described according to certain embodiments and generally associated methods, changes and variations of the embodiments and methods will be apparent to those skilled in the art. Therefore, the above description of exemplary embodiments does not limit this disclosure. Other changes, substitutions, and modifications are possible without departing from the spirit and scope of this disclosure. Furthermore, unless specifically stated otherwise, discussions using terms such as “analyze,” “determine,” “enable,” “identify,” and “modify” refer to the actions and processes of a computer system or similar electronic computing device that manipulate and transform data representing physical (e.g., electronic) quantities within the registers and memories of the computer system into other data representing physical quantities within the computer system's memory or other such information storage, transmission, or display devices.

[0035] It should be understood that the above description is intended to be illustrative and not restrictive. Many other embodiments will become apparent to those skilled in the art after reading and understanding the above description.

Claims

1. A method for processing medical images, the method comprising: Obtain one or more first medical images of a person and one or more second medical images of the person, wherein the one or more first medical images are associated with a first mammographic view of the person, and wherein the one or more second medical images are associated with a second mammographic view of the person; Determine a first set of features associated with the one or more first medical images; Determine a second set of features associated with the one or more second medical images; A third feature set is determined based on the first feature set and the second feature set, wherein the third feature set represents the mammographic features of the person indicated by the first mammographic view and the second mammographic view; and The presence or absence of medical abnormalities is predicted based at least on the third set of features. Determining a third feature set based on the first feature set and the second feature set includes: The features in the first feature set and the second feature set are fused into the corresponding specific view features representing the mammographic features of each view. Based on the fused specific view features, the cross-view features of the breast are determined to obtain the third feature set, wherein the cross-view features are features associated with multiple views.

2. The method according to claim 1, wherein, Predicting the presence or absence of the medical abnormality includes generating a first output image that includes a first indication of the medical abnormality.

3. The method according to claim 2, wherein, The first indication includes the boundary shape around the medical abnormality.

4. The method according to claim 2, wherein, The first indication includes a segmentation mask or heat map associated with the medical abnormality.

5. The method according to claim 2, wherein, Predicting the presence or absence of the medical abnormality also includes generating a second output image that includes a second indication of the medical abnormality, the first output image corresponding to the first mammographic view and the second output image corresponding to the second mammographic view.

6. The method according to claim 1, wherein, Predicting the presence or absence of the medical abnormality includes generating a classification label indicating the presence or absence of the medical abnormality.

7. The method according to claim 1, wherein, The one or more first medical images include a first digital breast tomography (DBT) image or a first full-field digital mammography (FFDM) image, and wherein the one or more second medical images include a second DBT image or a second FFDM image.

8. The method according to claim 1, wherein, The third feature set is determined using an artificial neural network (ANN), which includes a first encoding module and a second encoding module. The first encoding module is configured to determine the first feature set associated with the one or more first medical images or the second feature set associated with the one or more second medical images, and the second encoding module is configured to determine the third feature set.

9. The method according to claim 8, wherein, The ANN also includes a decoding module configured to predict the presence or absence of the medical abnormality based at least on the third feature set.

10. A non-transitory computer-readable medium comprising instructions that, when executed by a processor included in a computing device, cause the processor to: Obtain one or more first medical images of a person and one or more second medical images of the person, wherein, The one or more first medical images are associated with a first mammographic view of the person, and wherein the one or more second medical images are associated with a second mammographic view of the person; Determine a first set of features associated with the one or more first medical images; Determine a second set of features associated with the one or more second medical images; A third feature set is determined based on the first feature set and the second feature set, wherein the third feature set represents the mammographic features of the person indicated by the first mammographic view and the second mammographic view; and The presence or absence of medical abnormalities is predicted based at least on the third set of features. Determining a third feature set based on the first feature set and the second feature set includes: The features in the first feature set and the second feature set are fused into the corresponding specific view features representing the mammographic features of each view. Based on the fused specific view features, the cross-view features of the breast are determined to obtain the third feature set, wherein the cross-view features are features associated with multiple views.

Images