Deep learning technique for automated radiological image analysis and disease detection

Abstract

A real-time artificial intelligence (AI) framework is provided for the automated analysis of radiological images and detection of disease, such as extracapsular extension (ECE) in prostate cancer. The system includes a dual deep learning architecture comprising a first convolutional neural network (CNN) for identifying diagnostically relevant image slices from three-dimensional MRI data, and a second CNN for classifying disease presence based on those slices. A preprocessing pipeline standardizes and harmonizes image input, and cropping algorithms isolate the region of interest for enhanced model performance. This framework enables scalable, high-accuracy diagnosis across various imaging modalities including but not limited to MRI, CT, PET, ultrasound, and diverse disease types, improving clinical decision-making and supporting integration into real-time radiology workflows.

Description

DEEP LEARNING TECHNIQUE FOR AUTOMATEDRADIOLOGICAL IMAGE ANALYSIS AND DISEASE DETECTIONRELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Application Serial No. 63 / 643,711 filed May 7, 2024, entitled “DEEP LEARNING METHODS FORPROSTATE CANCER DIAGNOSIS,” the entirety of which is incorporated by reference herein.FIELD

[0002] The present concepts relate generally to artificial intelligence for medical imaging applications, and more specifically, to a real-time artificial intelligence (Al) framework for automating radiological image analysis, disease detection, and diagnostic decision supportBACKGROUND

[0003] Accurate preoperative staging of prostate cancer (PCa) plays a pivotal role in tailoring optimal treatment strategies, ensuring that patients neither undergo undertreatment nor overtreatment. Among the critical determinants of disease progression, the presence of extracapsular extension (ECE), identified as stage T3a or higher, emerges as a prominent factor, encompassing approximately one-third of newly diagnosed PCa cases. ECE detection refers to identifying when cancer cells break through the capsule of a gland or lymph node and extend into surrounding tissues, which is crucial for accurate cancer staging and treatment planning. ECE's significance lies inits close association with elevated rates of positive surgical margins and early biochemical recurrence following radical prostatectomy (RP), making it a key consideration in clinical decision-making. ECE has a significant prognostic value, decisively marking whether the cancer has infiltrated beyond the confines of the prostate gland. Many studies have proposed that not only the presence but also the amount of ECE is an independent predictive factor for biochemical recurrence (BCR). This further highlights the relevance of accurately predicting the presence of ECE before surgery.

[0004] While predictive nomograms based on clinicopathological attributes as well as magnetic resonance imaging (MRI) have been previously developed and validated, the intricate anatomy of the prostate and the subtle contrast between cancerous and healthy tissues present formidable challenges for ECE detection on MRI images.SUMMARY

[0005] The present disclosure relates to an artificial intelligence (Al) system for automated radiological image analysis and disease detection. The system leverages a dual deep learning (DL) pipeline comprising a first convolutional neural network (CNN) that filters relevant slices from a three-dimensional MRI scan and a second CNN that classifies the presence of disease, such as extracapsular extension (ECE) in prostate cancer. The preprocessing steps include image slicing, padding, harmonization, and cropping to focus analysis on a region of interest. The method is adaptable to various imaging modalities including CT, PET, and ultrasound, and it supports real-time processing for clinical deployment. Key innovations include the lightweight CNN architecture optimized for radiology, multislice analysis for improved predictionaccuracy, and an end-to-end software system deployable as a web application, implemented via executable software stored on non-transitory machine-readable media. The system can be applied across oncology, neurology, cardiology, and other specialties, offering significant improvements in diagnostic speed, accuracy, and workflow integration.

[0006] In one aspect, a real-time artificial intelligence (Al) framework is provided for the automated analysis of radiological images and detection of disease, such as extracapsular extension (ECE) in prostate cancer. The system includes a dual deep learning architecture comprising a first convolutional neural network (CNN) for identifying diagnostically relevant image slices from three-dimensional MRI data, and a second CNN for classifying disease presence based on those slices. A preprocessing pipeline standardizes and harmonizes image input, and cropping algorithms isolate the region of interest for enhanced model performance. This framework enables scalable, high-accuracy diagnosis across various imaging modalities and disease types, improving clinical decision-making and supporting integration into real-time radiology workflows.

[0007] In one aspect, an image processing system for diagnosing a medical condition comprises an input for receiving a plurality of images of an anatomical region from a medical instrument (MRI machine); a preprocessing system that standardizes the images to have common dimensions; a first deep learning model identifying and selecting diagnostically significant image slices of the standardized images; and a second deep learning model that evaluates the diagnostically significant image slices for abnormalities at the anatomical region.

[0008] In another aspect, an image processing system for diagnosing a possiblepreoperative staging of prostate cancer (PCa) comprises an input for receiving a plurality of images of a prostate gland from a medical instrument (MRI machine); a preprocessing system that standardizes the images to have common dimensions; a first deep learning model identifying and selecting diagnostically significant image slices of the standardized images; and a second deep learning model that evaluates the diagnostically significant image slices for a presence of extracapsular extension (ECE) at the prostate gland.

[0009] In another aspect, a method comprises a) uploading an image file, wherein the image file is a three-dimensional magnetic resonance image (MRI) that includes a prostate image of a prostate; b) multi-slicing the image to provide a plurality of two- dimensional sliced images; c) harmonizing each sliced image to provide respective a harmonized image with a standardized dimension using a script to pad the two-dimensional images; d) utilizing a first deep learning model with a first convolutional neural network (CNN) architecture to group the harmonized images into (1) distinct slice groups that display the prostate and (2) non-distinct slice groups that do not; e) cropping each harmonized image in the distinct slice group around a region of interest (RO I) that includes the prostate, thereby producing cropped images; and f) employing a second deep learning model that is a second CNN trained to classify the cropped images as either extracapsular extension positive (ECE+) or extracapsular extension negative (ECE-).

[0010] In some embodiments, the method further comprises quantifying a percent probability of ECE+ for each of the cropped images.

[0011] In some embodiments, the method further comprises quantifying an aggregate ECE+.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and further advantages of this invention may be better understood by referring to the following description in conjunction with the accompanying drawings, in which like numerals indicate like structural elements and features in various figures. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the invention. In the drawings:

[0013] FIG. 1 is a block diagram of a system for ECE detection, in accordance with some embodiments.

[0014] FIG. 2 is a flow diagram of a method for multi-modal deep learning model for ECE detection, in accordance with some embodiments.

[0015] FIG. 3 is a depicts a diagram of an artificial intelligence (Al) pipeline for ECE detection, in accordance with an example embodiment.

[0016] FIGs. 4 A and 4B are sequence diagrams illustrating two deep learning neural network models, respectively, each for image analysis, in accordance with some embodiments.

[0017] FIGs. 5A-5D are graphs illustrating results of a distribution across training, validation, and testing data sets and characteristics of images in a deep learning model, in accordance with some embodiments.

[0018] FIGs. 6A-6D are graphs illustrating performance results of the deep learning models of a detection system, in accordance with some embodiments.

[0019] FIGs. 7A-7F are graphs illustrating performance results of comparative machine learning (ML) models, in accordance with some embodiments.

[0020] FIG. 8A-8D are graphs illustrating performance comparisons for ECE detection in MRI scans, in accordance with some embodiments.

[0021] FIG. 9 is a set of screenshots of an output of an image processing system, in accordance with some embodiments.DETAILED DESCRIPTION

[0022] In the ever-evolving landscape of medical imaging and diagnosis, machine learning (ML) and deep learning (DL) algorithms have emerged as revolutionary tools with widespread applications. Within this context, DL models, which are a subset of ML models, have demonstrated remarkable capabilities in deciphering intricate patterns and relationships within medical images. This extends to the domain of PCa diagnosis, where DL algorithms have shown promise. Also, with the lack of a truly objective method of assessment to evaluate the presence of PCa on MRI, one must turn to indirect signs to confirm or deny the likelihood of a cancer diagnosis.

[0023] Traditionally, an ECE lesion on a prostate or area of concern captured on an MRI or other image that shows evidence of extracapsular extension (ECE), may indicate that cancer has grown beyond the prostate capsule. PLRADS (Prostate Imaging Reporting and Data System) is used to assess the likelihood of clinically significant cancer, and lesions are assigned a score, with higher scores indicating a greater likelihood of cancer. However, identifying ECE as well as lesion identification of MRI images using PLRADS has been a time-intensive process heavily reliant on the expertise of radiologists. However, the advent of artificial intelligence has opened new avenues for automating this intricate task. By training convolutional neural networks (CNNs) on extensive datasets of annotated MRI scans, these algorithms can learn to recognize subtlevisual cues indicative of ECE. However, current Al solutions are limited to diagnosing specific cancer types, single-slice image analysis, or request manual feature extraction, e.g., using radiomics-based models. Other ALassisted technologies are incapable of operating with respect to specific applications. There is a desire for an Al solution to optimize the delicate balance between accuracy and efficiency, which is not offered by current Al technology.

[0024] In brief overview, embodiments of the present inventive concept provide a comprehensive, fully automated artificial intelligence (Al) pipeline constructed for ECE detection in diseases such as prostate cancer and can be performed on T2-weighted MRI scans or the like. Features of a system incorporating the Al pipeline can be implemented via executable software stored on non-transitory machine-readable media and executed by special-purpose computer hardware to perform preprocessing operations such as image harmonization prior to receipt by deep learning (DL) models for performing detailed analysis of the MRI scans, or more specifically, two-dimension (2D) slices in various planes from three-dimensional images 3D representing cross-sectional views of the anatomy such as the prostate, captured by the MRI images. Each two-dimensional “slice” can be taken from a specific plane, e.g., axial, coronal, sagittal, and when combined, these slices provide a three-dimensional representation of the scanned area. By leveraging the capabilities of two advanced deep learning (DL) models, a systematic framework is offered for the processing and refinement of images to predict ECE in PCa for precise medical image analysis. A first DL model identifies relevant MRI images, or slices, for training, and the other model assesses the likelihood of ECE in patients, thereby achieving high accuracy. For example, an Al pipeline can be constructed andarranged for clinical diagnosis and staging by improving the accuracy of preoperative staging of PCa by reliably detecting ECE, which is a critical factor in determining the aggressiveness of the disease. This process extends from the radiologist's bench, through a series of sophisticated analytical stages, to the generation of outcomes that are easily interpretable by medical professionals, to streamline and enhance the diagnostic workflow. In some embodiments, the system includes a web application, for example, shown in FIG. 9 that can display an output as a user-friendly user interface for use as a diagnostics and treatment planning tool. Designed for clinical use, the application promises to facilitate the rapid and reliable identification of ECE across diverse MRI datasets. By improving both the accuracy and consistency of diagnoses, this approach early diagnostic intervention and supports more precise treatment planning, thereby making a significant contribution to the advancement of patient care.

[0025] Conventional methods rely heavily on the subjective interpretation of MRI scans by experienced radiologists, which can vary in accuracy. The systems and methods according to embodiments of the present inventive concept, on the other hand, standardizes ECE detection, reduces variability in diagnosis, and supports higher diagnostic confidence, leading to better patient management and outcomes.

[0026] The staging above can assist in tailoring optimal treatment strategies, ensuring patients receive appropriate intervention without the risks of undertreatment or overtreatment. In some embodiments, an Al pipeline is constructed and arranged for treatment planning. By providing precise information about the presence of ECE, the Al pipeline, illustrated for example in FIG. 3, can aid surgeons and oncologists in planning more effective surgical interventions, such as deciding on the extent of tissue removalduring radical prostatectomy to achieve clean surgical margins while preserving as much healthy tissue as possible. In some embodiments, the Al pipeline is constructed and arranged as an educational tool. Here, the Al pipeline implemented via executable software stored on non-transitory machine-readable media and executed by specialpurpose computer hardware can serve as an educational resource for medical professionals, enhancing their understanding of PCa imaging and diagnosis through exposure to advanced Al diagnostics.

[0027] Additional applications include (1) expansion to other cancers: The technology behind the Al pipeline has the potential to be adapted for the detection and analysis of other types of MR images (e.g. T1W, DWI) and cancers such as breast cancer where imaging plays a crucial role in diagnosis and staging, thus broadening its impact in oncology. (2) Integration with Radiomics: Future developments could integrate radiomics features, which provide quantitative data from images that humans cannot perceive, enhancing the model's diagnostic accuracy further. (3) Real-time Diagnostic Support: Al pipeline could be developed to provide real-time feedback during medical imaging sessions, offering immediate insights to radiologists and reducing the time between imaging and diagnosis.

[0028] The Al pipeline introduces several advantages and novel features over existing technologies in the detection of ECE in PCa through MRI scans. On advantage lies in the implementation of two specialized deep learning models. The first model efficiently identifies the most informative MRI slices for analysis, while the second model focuses on accurately predicting ECE presence. This dual-model approach enhances accuracy and specificity compared to single algorithm systems.

[0029] In some embodiments, the Al pipeline includes a user-friendly interface, and in doing so is constructed as a web application, making it easily accessible to clinicians and radiologists without the need for advanced technical knowledge. This enhances usability in clinical settings, and provides high diagnostic accuracy as shown in in clinical trials, with area under the curve (AUC) metrics of 0.92 for the first model and 0.88 for the second model (shown in FIGs. 6A and 6C, respectively) significantly outperforming traditional diagnostic methods.

[0030] By leveraging a large dataset annotated by experts, the Al pipeline offers robust performance that is data-driven, providing a strong empirical basis for its diagnostic predictions. In some embodiments, the Al pipeline can perform real-time analysis during MRI procedures, providing immediate diagnostic insights.

[0031] Accordingly, the dual deep learning model approach according to embodiments is tailored specifically for PCa and ECE detection, user-friendly interface, and potential for real-time image analysis set it apart from existing technologies, which often lack specificity, adaptability, or ease of use in clinical settings. Although embodiments of the present inventive concept describe a versatile Al framework for ECE detection for prostate cancer, applications are not limited thereto. The system can be retrained and subsequently deployed to analyze additional medical imaging task across oncology, neurology, cardiology, musculoskeletal disorders, pulmonary diseases, and the like, and can analyze radiological imaging datasets corresponding to these multiple medical conditions.

[0032] FIG. 1 is a block diagram of a system 100 for ECE detection, in accordance with some embodiments. As shown, the system 100 in some embodimentsincludes a medical instrument such as an MRI machine 110, an imaging processing system 120, and a display apparatus 140.

[0033] The MRI machine 110 generates a combination of magnetic fields and radio waves to produce three dimensional anatomical images of a region of interest of a human body, for example, a prostate gland but not limited thereto. In some embodiments, the images are MRI scans, for example, T2-weighted MRI scans or preoperative MRI scans of PCa patients. In some embodiments, the images are MR images or other three-dimensional images used for detecting cancers such as breast cancer where imaging plays a crucial role in diagnosis and staging and used in oncology. In other embodiments, instead of MRI images, other radiology apparatuses may provide images, such as computed tomography (CT) scans, Positron Emission Tomography (PET) scans, X-rays, ultrasound images, and / or other various imaging modalities.

[0034] The image processing system 120 includes a slicing module 121, a converting module 122, a padding module 123, a harmonizer module 124, a first deep learning model 126, a cropping module 128, and a second deep learning model 130.

[0035] The slicing module 121 can extract and view two-dimension (2D) slices in various planes from a three-dimensional images 3D for multi-slicing analysis. In some embodiments, the slicing module 121 can be part of an input module that receives images from an image generating source, for example, the MRI machine 110, for processing. The input module may include a computer-readable storage device constructed and arranged to store the received images, for example, a computer hard drive, solid state device, integrated circuit memory, cloud storage device, and so on.

[0036] The 2D slices produced by the slicing module 121 may represent cross-sectional views of the body, and they are typically organized in a series of 2D images, each capturing a specific slice of the three-dimensional anatomy. These slices can be aligned in different orientations, such as axial, coronal, or sagittal planes. When analyzing images from multiple slices, the padding module 123 and harmonizer module 124 can ensure consistency in how the data is processed when there arc differences in slice alignment, resolution, or scanning conditions. For example, the slices may have discrepancies in terms of resolution, intensity, or size due to various factors like hardware limitations, patient positioning, or motion artifacts. Padding harmonization is an approach to address these issues.

[0037] In some embodiments, the padding module 123 can add pixels (or voxels in 3D data) around the edges of the MRI slices to ensure that the slices are of uniform size and can be consistently processed. This can be used to make each slice's dimension consistent, especially when the slices are uneven due to various factors. This preprocessing operation provides a structural adjustment that ensures that all 2D image slices are the same dimensions e.g., square dimensions of 512x512 pixels by adding empty pixels around smaller images. In some embodiments, a separate script, e.g., a Python script or the like, is used for the padding module 123 to perform such padding operations.

[0038] The harmonizer module 125 performs another preprocessing operation, namely, to standardizing image characteristics like intensity, contrast, and resolution across images to ensure consistency — especially when MRI images come from different machines or protocols. This could mean adjusting brightness / contrast or compensating for slight misalignments between slices. The abovementioned registration step can beused for aligning the slices to a common reference or coordinate system. This can correct for slight misalignments between slices and ensure that spatial relationships are maintained. The normalization step, e.g., performed by the harmonizer module 124, can be used for adjusting intensity or color levels across slices to ensure that there are no biases due to different imaging conditions (lighting, exposure, etc.). Interpolation techniques can be used to create a uniform dataset if slices arc spaced unevenly or at different resolutions.

[0039] The converting module 122, padding module 123, and harmonizer module 124 are constructed and arranged to standardize the dimensions, resolution, and / or preprocessing steps across multiple MRI images to make them comparable and suitable for analysis by the machine learning models of the system 120. The converting module 122, padding module 123 and harmonizer module 124 may be generated by a set of Python scripts or the like for detailed preprocessing. A first specialized script may be used by the converting module 122 to convert images into a consistent format. A second specialized script may be used by the padding module 123 to pad the images, for example, adjust the dimensions of an MRI image to a consistent size. The padding module 123 ensures all image slices from the converting module 122 are the same size.

[0040] The harmonizer module 124 may make all image slices consistent in terms of certain factors, including resolution, intensity normalization, and alignment, for example, by performing color correction, contrast adjustment and / or normalization to standardize pixel values across the images, or the like. The harmonizer module 124 may apply techniques such as resampling, intensity normalization, and so on. The resampling step may include resizing the received MRI images to a common resolution, e.g., allimages should have the same voxel size. Intensity normalization techniques can be applied to standardize the image intensity across different datasets, ensuring consistency in signal. In addition, but not limited, other techniques may include aligning all images to a common coordinate system (e.g., using affine or non-rigid registration) to ensure consistency in the spatial structure across all images and adjusting the contrast of the images to match them or standardizing them based on a reference image. Other preprocessing may include smoothing and noise reduction, bias field correction, and segmentation for isolating the regions of interest before further processing by the machine learning models.

[0041] In sum, each sliced image can be harmonized to provide a harmonized image with a standardized dimension using a script to pad the two-dimensional images. This may refer to a harmonization step that includes padding as part of standardizing image dimensions. Padding is one sub-process used during harmonization, where a script (e.g., written in Python) adds pixels around the edges of a 2D image slice so that all images match a uniform dimension (e.g., 512x512 pixels). Harmonization, more broadly, refers to a set of preprocessing techniques applied to MRI images to standardize characteristics across a dataset. These can include padding, but also intensity normalization (e.g., scaling pixel values), resampling (to unify resolution or voxel spacing), and / or alignment or registration (ensuring anatomical landmarks are aligned across slices). As previously mentioned, the harmonization includes the padding step as the method used to achieve standardized dimensions. However, the script doesn't pad after harmonization, instead, padding is part of the harmonization operation.

[0042] The first DL model 126 efficiently identifies the most informative andrelevant MR1 slices for training and analysis, in particular, assesses the image quality of the most informative and diagnostically significant MRI slices for a given patient. More specifically, the first DL model 126 identifies slides that include the presence of the gland, organ, or other anatomical location of interest. For example, the first DL model 126 is tasked with identifying multiparametric MRI (mpMRI) images where the prostate gland or anatomical region of interest is discernible. These identified images then serve as the training dataset for the second model 130. The first DL model 126 could aid in detecting slices and harmonizing them before analysis. This multislice approach not only increases the volume of data (breadth) but also enhances the granularity of information captured from each patient (detail). By analyzing several contiguous slices, a more comprehensive representation of the anatomical structures is achieved, which is crucial for accurate diagnosis and analysis. This method significantly enriches the received dataset, enhancing both its breadth and detail, thereby providing a more robust basis for the deep learning models, namely, the two convolutional neural networks of the DL models 126, 130. In some embodiments, the primary dataset of MRI images or the like for processing by the two DL models may include two types of images classified by expert radiologists as 'Distinct slices', which clearly show the prostate gland, and 'Non- Distinct slices', which depict the pelvic anatomy, including muscles and other organs, but do not provide a clear view of the prostate gland itself. The first DL model 126 automatically selects all the distinct images from the original dataset of MRI images provided by the MRI machine 110. In embodiments where other image types are processed, other medical instruments may equally apply.

[0043] The cropping module 128 can crop each harmonized image in the distinctslice group grouped by the first deep learning model with the first CNN around a region of interest (RO I) that includes the region of the anatomy, e.g., prostate, thereby producing cropped images. As described herein, this preprocessing step is critical to ensure that on optimal focus is made on the relevant anatomical region. For example, the images around the prostate gland can be cropped to ensure the algorithm's analysis is focused on the presence of the tumor either within or immediately adjacent to the prostate gland or seminal vesicle, while avoiding analysis of surrounding structures such as the bladder, rectum, levator ani muscles, obturator intemus muscles, and pelvic bone. To accomplish this targeted image cropping, a specialized script, delineated and described herein.

[0044] The second deep learning model 130 like the first deep learning model 130 may include a CNN type deep learning model. However, the second deep learning model 130 has a classification model for classifying preprocessed images output from the cropping module 128 as either ECE-positive or ECE-negative. This multifaceted approach using two different models 126, 130 can produce results precisely tailored to investigate the intricate dynamics and patterns associated with PCa and ECE.

[0045] The display 140 includes a user interface produced by executable software implemented as a web application and stored on non-transitory machine -readable media, making it easily accessible to clinicians and radiologists without the need for advanced technical knowledge. This enhances usability in clinical settings.

[0046] FIG. 2 is a workflow diagram of a method 200 for multi-modal deep learning model for ECE detection, in accordance with some embodiments. In describing the method 200, reference is made to elements of FIG. 1. In some embodiments, the method 200 can be applied to predict ECE in PCa but not limited thereto.

[0047] At step 210, at least one medical image input is received. In some embodiments, the medical image input is a dataset of MR images, e.g., multiparametric MRI (mpMRI) images but not limited thereto.

[0048] At step 220, the images received at step 210 are preprocessed. In some embodiments, one or more of the modules 121 (slicing module), 122 (converting module), 123 (padding module), and 124 (harmonizer module) of FIG. 1 contribute to the preprocessing at step 220. Preprocessing sub-steps may include but not be limited to the slicing module 121 producing 2D images. Preprocessing examples may include but not be limited to the slicing module 121 generating 2D images from a 3D MRI scan, the converting module 122 converting the 2D image slices from a Digital Imaging and Communications in Medicine (DICOM) format to a Portable Network Graphics (PNG) format or the like, the padding module 123 padding the PNG-formatted images to achieve square dimensions, and the harmonizer further modifying the images to blend seamlessly with each other. Preprocessing MRI images across diverse datasets facilitates rapid and reliable identification of ECE across the diverse MRI datasets, improving both accuracy and consistency in diagnoses.

[0049] At step 230, a first DL model is applied to the distinct slices of the images, i.e., clearly showing the prostate gland or other organ or region of the anatomy of interest preprocessed, e.g., padded to achieve square dimensions for image harmonization, in step 220 and provided as input for training the first DL model. The first DL model selects distinct slices from the received input of image slices based on the presence of the prostate gland or region of the anatomy of interest.

[0050] At step 240, a second DL model is applied to the distinct image slicesselected by the first DL model and classifies the images as either ECE-positive or ECE- negative.

[0051] FIG. 3 is a sequence diagram of an artificial intelligence (Al) pipeline 300 for ECE detection, in accordance with an example embodiment. In describing FIG. 3, reference is made to elements of the ECE detection system 100 of FIG. 1. Although the pipeline 300 performs predictive ECE status techniques, the pipeline 300 may equally apply with respect to techniques for performing other classification tasks in the medical field.

[0052] In cases where the input images are in a DICOM format or the like, a set of input images, e.g., MRI images, undergo one or more preprocessing operations (302). In some embodiments, the preprocessing operations may include converting and normalizing the input images, applying enhanced color scaling, and padding the images to square dimensions for compatibility with CNN-based processing or the like. For example, at step 302, the converting module 122 can convert (301) one or more input images from the DICOM format to a PNG format using a Python script or the like, which is suitable for subsequent processing by the DL models 126, 130. Another preprocessing operation (302) may include the padding module 123 modifying the images to achieve standardized dimensions for image harmonization. The conversion and padding steps (302) can be performed in sequence using specialized scripts or the like. In response the padding step, the images have a uniform resolution, e.g., 512 / 512 pixels, thereby maintaining the original resolution and quality of the original images. In some embodiments, these preprocessing steps are part of a normalization process that may include a harmonization step applied to ensure consistency across the dataset.

[0053] As described above, the conversion and padding steps are sequential. First, MRI scans (which may be 3D volumes in DICOM format) are sliced into 2D images, if not already received that way. These 2D slices are then converted to PNG, followed by padding to achieve uniform dimensions. Accordingly, the processing order starts with a 3D scan (if needed), followed by slicing, conversion to PNG or the like, padding, and finally harmonization. Although pre-sliced 2D inputs are described in some embodiments, in other embodiments, full 3D-to-2D slicing may be performed, e.g., from DICOM when necessary.

[0054] In some embodiments, scans, e.g., MRI image scans as 3D volumes may be in a DICOM format or the like and slices the 3D DICOM into 2D images in preparation of the preprocessing operations. For example, the 2D slices are then converted to PNG, followed by padding to achieve uniform dimensions.

[0055] The preprocessed, or harmonized, image slices are applied to the first DL model 126, which can identify and select (304) distinct slices that include a region of interest (ROI), e.g., presence of the prostate gland. The first DL Model 126 can distinguish distinct slices, which clearly show the prostate gland, from non-distinct slices, which depict the pelvic anatomy, including muscles and other organs, but do not provide a clear view of the prostate gland itself. In some embodiments, the first DL model 126 can apply a multislice method to processes multiple MRI slices per patient automatically. This multislice approach not only increases the volume of data (breadth) but also enhances the granularity of information captured from each patient (detail). By analyzing several contiguous slices, a more comprehensive representation of the anatomical structures is gained, which is crucial for accurate diagnosis and analysis. This methodsignificantly enriches the dataset of received images of the anatomical region of interest, enhancing both its breadth and detail, thereby providing a more robust basis for the artificial intelligence models.

[0056] Next, the slices selected (304) can undergo a preprocessing step to ensure an optimal focus on the relevant anatomical region. Specifically, selected slices around the prostate gland are cropped (306) to ensure the algorithm's analysis is focused on the presence of the tumor either within or immediately adjacent to the prostate gland or seminal vesicle rather than invading the bladder, rectum, levator ani and obturator intemus muscles, and the pelvic bone. To accomplish this automatic targeted image cropping, a specialized script is provided, which can crop (306) the images around the region of interest (ROI) while minimizing or eliminating irrelevant surrounding structures. It is imperative to crop (306) the images around the prostate gland after the images are harmonized (302) to ensure the algorithm's analysis is focused on the presence of the tumor either within or immediately adjacent to the prostate gland or seminal vesicle rather than invading the bladder, rectum, levator ani and obturator intemus muscles, and the pelvic bone.

[0057] At step 308, pathology-confirmed labels can be assigned to the preoperative MRI image slices based on post-surgery histopathological findings. This can be part of a pathology-radiology fusion operation. For example, the labels can be assigned as ECE-positive or ECE-negative. After preprocessing and slice selection, each MRI dataset is matched to the corresponding patient's post-surgery histopathology report. The pathology findings, which are considered the ground truth, confirm whether extracapsular extension (ECE) was present or absent. Based on this confirmation, thepreoperative MRI slices are assigned a binary label: 'ECE-positive' (ECE+) if ECE is present, or 'ECE-negative' (ECE-) if absent. This process, termed pathology-radiology fusion, enables supervised learning by providing reliable target labels for training the second deep learning model. By aligning the radiological imaging data with pathology- confirmed outcomes, the system ensures accurate and clinically meaningful classification during model development.

[0058] Upon completion of the preprocessing steps, the images are aptly prepared for the final analytical phase — classification (310) by the second DL model 130, which in some embodiments discerns between ECE+ and ECE- cases, and in other embodiments performs other classification tasks as required. This methodological approach ensures that the dataset is not only multifaceted but also precisely tailored to investigate the intricate dynamics and patterns associated with PCa and ECE.

[0059] The output of the classification step (310) may be processed in an Al clinical data fusion step (312). Here, a multimodal late fusion operation may be performed by integrating the Al-derived imaging features produced by the second model 130 with clinical data such as ISUP grade, patient age, and so on. The machine learning models can be used to enhance ECE as described above.

[0060] FIGs. 4A and 4B are diagrams illustrating two deep learning neural network models for image analysis, in accordance with some embodiments.

[0061] In some embodiments, the first and second deep learning models 126, 130, respectively, described in FIGs. 1-3 may apply to a multi-convolutional neural network (multi-CNN) model, which is a critical element of the inventive methodology for detecting ECE in PCa, in accordance with some embodiments. The multi-CNN modelstands out due to several key features that collectively bolster its capability in the complex task of identifying ECE from MRI scans of PCa patients. The underlying structure of the multi-CNN model comprises two sequential CNN algorithms, each designed with specific, interconnected functions in the ECE detection workflow. In some embodiments, both CNNs are trained from scratch, eschewing the use of any pre-trained models in this research. Initially, the algorithm embodied in the first model 126 is responsible for identifying and selecting the most informative and diagnostically significant MRI slices for each patient. Following this, the algorithm in the second model 130 utilizes these chosen slices to categorize patients into positive or negative ECE groups. This classification is based on MRI images that have been labeled in accordance with ground truth post-surgery pathology results.

[0062] In some embodiments, the system and method for detecting ECE within MRI scans includes the training of each algorithm from scratch, a strategy that bypasses the limitations tied to the pre-existing knowledge found in transfer learning. This approach grants the models 126, 130 a deeper and more nuanced comprehension of a specific dataset, enabling them to adjust their parameters with enhanced precision for the distinct features of MRI images pertinent to ECE detection. By designing the models 126, 130 specifically for this task, one inherently avoids the common issue of overfitting, which frequently affects complex, pre-trained models when they are repurposed for specialized applications.

[0063] As shown in FIGs. 4A and 4B, illustrations of the two deep learning models 126, 130 applying CNN architectures 400, 450, respectively, are more efficient and effective compared to more deeply layered counterparts. The architectural diagrams400, 450 are identical except for the first diagram 400 corresponding to the first deep learning model 126 of FIG. 1, step 230 of FIG. 2, and step 304 of FIG. 3 and the second diagram 450 corresponding to the second deep learning model 130 of FIG. 1, step 240 of FIG. 2, and step 310 of FIG. 3. This not only minimizes the risk of overfitting but also improves the models' ability to predict using new, unseen MRI data effectively. Furthermore, the streamlined nature of the models aligns with the practical requirements of clinical environments. They strike a careful balance between computational efficiency and diagnostic precision, making them highly suitable for quick adoption into clinical workflows. The models' reduced processing demands enable swifter, more efficient analyses without compromising on accuracy, addressing the critical needs of healthcare settings where both time and precision are of utmost importance.

[0064] In some embodiments, the architectural diagrams 400, 450 illustrate a CNN architecture tailored for efficient image classification, utilizing depthwise separable convolutions to reduce computational complexity while maintaining performance. In some embodiments, the CNN of the first model 126 comprises three primary depthwise separable convolutional layers 402A, 402B, 402C with increasing filter sizes utilizing depthwise layers (Conv2D), each followed by batch normalization (BatchNorm2D), ReLU activation (ReLU), and max pooling (MaxPool2D) features, culminating in a fully connected classification layer 402C(Linear) in FIG. 4A, the entire model is designed for compatibility with both standard CPU and Compute Unified Device Architecture (CUDA)-enabled graphics processing unit (GPU) environments, enabling flexible deployment and efficient processing across diverse hardware platforms.

[0065] In some embodiments, the CNN 450 of the second model 130, similar toCNN 400, comprises a set of convolutional layers providing depthwise and pointwise convolutions, each followed by batch normalization and a ReLU activation function followed by dense layers with dropout regularization. Each convolution block is accompanied by a max pooling layer depthwise Conv2D layers, namely, four primary depthwise separable convolutional layers 452A-D. After the final pooling layer 460 , the network transitions to a dense segment 460 consisting of two fully connected layers separated by a dropout layer set at 60% to mitigate overfitting. The flattened output from the pooling layers is passed through these layers, with the final layer outputting a probability distribution across the predefined number of classes.

[0066] In some embodiments, the CNN architecture 450 of the second DL model 130 shown in FIG. 4B is facilitated by a robust preprocessing pipeline, for example, using a transform module provided by a domain-specific library such as Python’s Torchvision but not limited thereto. In some embodiments, during preprocessing, image inputs are resized, for example, to 512x512 pixels, to enhance detail recognition, randomly flipped horizontally to augment the dataset and improve generalization, and normalized to scale pixel values between -1 and 1. The training, validation, and test datasets are loaded using a data loader for handling large datasets efficiently, and providing features like batching, shuffling, and parallel loading of data. For example, PyTorch's DataLoader but not limited thereto, with shuffled batches of 64 and 32 images for training and validation, respectively, can be used promote model robustness. In this experiment example, the optimization process employs an RMSprop optimizer configured with a slightly elevated initial learning rate of 0.00001, an alpha of 0.9 for smoothing, and momentum set at 0.5 to accelerate convergence in the appropriatedirection of the gradients. Weight decay is also used to reduce overfitting by penalizing large weights. The loss function employed is the CrossEntropyLoss, which is particularly effective for classification tasks involving multiple classes by combining LogSoftmax and NLLLoss in a single criterion.

[0067] As shown in FIGs. 5A-5D, the distribution of data across training, validation, and testing sets is meticulously balanced concerning patient age, preoperative biopsy pathological grade, and ECE outcomes. FIGs. 5A-5D provide a detailed overview of the dataset, which comprises a total of 4200 images from 701 patients allocated for training purposes, 1340 images from 200 patients designated for validation, and 622 images from 100 patients intended for testing.

[0068] Importantly, it highlights the mean age of patients within each subset — 64.0 years for the training group, 65.41 years for the validation group, and 63.03 years for the testing group — thereby showcasing a consistent age distribution across all datasets.Additionally, a visual depiction of age distribution curves for each dataset set is presented, emphasizing the homogeneity of this demographic factor. FIGs. 5C and 4D elucidate the distribution of ECE status according to biopsy grade groups via a mosaic plot, thus confirming the balanced representation of various pathological grades and ECE outcomes across the study cohorts. The inclusion of preoperative biopsy grade and patient age as clinical parameters aims to enhance the predictive accuracy of a decision tree model, enabling a multimodal approach to estimating the likelihood of ECE.

[0069] To explore the relationship between patient age and preoperative International Society of Urological Pathology (ISUP) grade group, embodiments of the system used for generating the data shown in FIGs. 5 A and 5B includes a combination ofmachine learning (ML) and deep learning (DL) processes. The graphical results shown are produced from the ML processes comprising Decision Tree Classifier (DTC), Random Forest, and XGBoost algorithms implemented via the Scikit-learn library in Python for predicting ECE using the ISUP grade group and patient age across a cohort of patients. Complementing these, the DL models are developed using a framework such as Pytouch or the like to harness advanced computational techniques. The framework may include a dataloader that loads and processes the training, validation, and test datasets shown in FIGs. 5A and 5B.

[0070] Before proceeding with the analysis, the dataset is prepared by converting categorical variables into dummy variables. This transformation, alongside the conversion of the ECE results into a binary outcome, allowed for seamless integration and analysis of both categorical and continuous variables. The DL models underwent training and evaluation on a T4 GPU, chosen for its efficiency and capability to handle complex computations inherent in DL tasks.

[0071] The data was partitioned into a training set and a testing set following a 70 / 30 ratio to ensure a balanced representation of outcomes. The models were configured with a maximum depth of four to maintain generalizability and prevent overfitting. The decision structure export displays nodes rounded for clarity, with the tree proportionally colored to reflect class distribution at each decision juncture.

[0072] With further regard to FIGs. 5A-5D, comprehensive statistical analyses were conducted, including Specificity, Sensitivity, and Accuracy (ACC) evaluations, along with Receiver Operating Characteristic (ROC) curve analysis to calculate the AreasUnder the Curve (AUCs). Comparisons against radiologists with varying expertise levelsin prostate mpMRI readings were made using the Jaccard index and Cohen's Kappa.

[0073] FIGs. 6A-6D illustrate the proficiency of the combination of first and second deep learning models illustrated in FIGs. 4A and 4B ProSliceFinder and ExCapNet models in identifying relevant scans and accurately predicting ECE. The second DL model, in particular is constructed and arranged to analyze individual image slices, exhibited impressive performance metrics on a blind test set comprising 467 MRI slices. As shown in FIG. 6A, this Areas Under the Curve (AUC) is 0.92, a sensitivity of 0.91, a specificity of 0.80, and an overall ACC of 0.85, thereby effectively identifying slices depicting the prostate gland using T2W imaging.

[0074] FIG. 6B illustrates a confusion matrix that corresponds to the performance of the first deep learning model, which is trained to classify MRI slices as either distinctive or non-distinctive. Distinctive MRI slices visibly include the prostate gland (anatomically informative for diagnosis). Non-distinctive MRI slices on the other hand do not clearly show the prostate (e.g., muscle tissue, surrounding pelvic areas). True Positive (Top-left: 183-Correctly predicted distinctive slices), False Negative (Top-right: 47-Slices that were distinctive, but the model predicted them as non-distinctive), False Positive (Bottom-left: 21 -Slices that were non-distinctive, but the model predicted them as distinctive), True Negative (Bottom-right: 216-Correctly predicted non-distinctive slices). FIG. 6B illustrates that the model in accordance with embodiments of the inventive concept performs well at identifying relevant (distinctive) slices to pass on to the second model 130. The high specificity ensures that non-informative slices are reliably excluded, which streamlines downstream classification and improves diagnostic efficiency.

[0075] Complementing this, the second model 130, which assesses data at the patient level, garnered as shown in FIG. 6C as an AUC of 0.88, Sensitivity of 0.84,Specificity of 0.80, and ACC of 0.82. It capitalizes on the slice-level insights provided by the second model 130 to make informed patient-level ECE predictions. This analysis was also performed on a blind test set, featuring 100 randomly selected patients, shown in FIGs. 6C and 6D, which include graphs illustrated the performance by the second model 130 for detecting distinct, i.e., visible prostate gland slices and ECDE, respectively, using ROC curves (FIG. 6C) and a confusion matrix (FIG. 6D).

[0076] The confusion matrix shown in FIG. 6D reflects the performance of the second deep learning model 130 in predicting extracapsular extension (ECE) at the patient level: Along the Y-axis are True Labels, which represent the ground truth classification of each patient, based on post-surgical pathology: “ECE Positive” = patients confirmed to have extracapsular extension. “ECE Negative” = patients without ECE.

[0077] Along the X-axis are Predicted Labels. These are the classifications made by the Al model. True Negative (Top-left: 40-The model correctly predicted 40 patients as ECE-negative who were indeed ECE-negative.), False Positive (Top-right: 10- The model predicted ECE-positive, but pathology confirmed these 10 patients did not have ECE), False Negative (Bottom-left: 8-The model predicted ECE-negative, but pathology confirmed these 8 patients did have ECE), True Positive (Bottom -right: 42-The model correctly predicted 42 patients as ECE-positive who were confirmed ECE-positive by pathology). These results indicate that the second DL model demonstrates strong predictive performance and maintains a good balance between sensitivity and specificity— an important consideration in preoperative cancer staging.

[0078] As shown in FIGs. 7A-7B, a comparative analysis is shown of three ML models — Decision Tree, Random Forest, and XGBoost — to assess their efficacy in predicting ECE using the ISUP grade group and patient age across a dataset of 1001 patients. In FIGs. 7C-7F, the comparison of the models’ predictive accuracy is graphically show within a subset of 100 randomly selected test patients, distinguished by the absence of Al-derived features (FIGs. 7C and 7D) and presence of Al-derived features (FIGs. 7E and 7F). FIGs. 7B, 7D, and 7F enumerate the key features determined by each algorithm to be most critical for predictive performance, segmented by model and dataset.

[0079] For the larger cohort of 1001 patients shown in FIGs. 7A and 7B, the resulting findings revealed moderate predictive capabilities across the models, with the Decision Tree achieving an ACC of 0.64 and an AUC of 0.65, Random Forest with an ACC of 0.62 and AUC of 0.65, and XGBoost leading slightly with an ACC of 0.63 and an AUC of 0.68. This initial analysis highlighted XGBoost’ s marginal superiority in AUC, suggesting nuanced variations in model performance.

[0080] Upon excluding Al-derived features in the subsequent 100-patient cohort, shown in FIGs. 7C-7F, a noticeable decrement in performance metrics was observed across all models, underscoring the significant role Al features play in enhancing predictive accuracy. Conversely, the integration of Al features not only recuperated but also enhanced the performance of the models. Specifically, the Decision Tree, Random Forest, and XGBoost models showed a significant increase, with their AUC values rising to 0.81, 0.81, and 0.78, respectively. Additionally, all three models achieved the sameACC of 0.73. This improvement accentuates the additive value of Al insights to the predictive analytics framework.

[0081] FIGs. 7A-7F also illustrate a pivotal shift toward Al predictions as a significant determinant of ECE, surpassing traditional clinical markers such as patient age and ISUP grade. This evolution in feature importance rankings evidences the potent synergy between clinical attributes and Al-derived features, fostering a refined approach to patient stratification.

[0082] The empirical evidence strongly advocates for the amalgamation of traditional clinical features with Al-driven insights. This integration not only enriches the predictive landscape but also significantly elevates the performance of ML models within clinical settings, paving the way for more informed and nuanced patient care strategies.

[0083] FIG. 8A-8D are graphs illustrating performance comparisons for ECE detection in MRI scans, in accordance with some embodiments. In particular, FIGs. 8A- 8D include results from a comparison between T2W MRI sequences and whole sequences of MRI data, including Tl-weighted (T1W), axial, and sagittal views. This comprehensive dataset provided them with a broader diagnostic context. This permits a comparison of the diagnostic accuracy of radiologists and Al against the ground truth from pathology reports.

[0084] In particular, the radiologists' diagnostic performance quantified using AUC and ACC, recorded values of 0.61 , 0.65, and 0.54, respectively, with the formal radiology report yielding an ACC of 0.58 (shown in FIGs. 8A and 8B). These figures suggest that determining ECE status solely through imaging remains challenging, even for seasoned practitioners. In contrast, the ECE detection system in accordance withembodiments demonstrates superior classification capabilities with an ACC of 0.82. Its agreement with the gold-standard pathology results, measured by Cohen’s kappa score, was Substantial (kappa = 0.64), notably higher than that of the individual radiologists (kappa = 0.22, 0.30, and 0.08) and the radiology report (kappa = 0.16), indicating a significant advancement in reliability and accuracy, shown in FIG. 8A.

[0085] To visualize these findings, FIG. 8C employs a circular heatmap, and FIG. 8D shows a dendrogram. The heatmap of FIG. 8C illustrates the degrees of concordance and discordance among the assessments by the radiologists, the formal radiology report, and the confirmed pathology cases, highlighting the spectrum of diagnostic alignment. Meanwhile, the dendrogram shown in FIG. 8C, constructed using the Jaccard distance index, clusters the radiologists based on the similarity of their diagnostic decisions and contrasts these with AutoRadAl’s outcomes. Jaccard distances of 0.57, 0.69, 0.68, and 0.79 were observed among the radiologists, with the Al system showing a closer alignment to the ground truth (Jaccard distance = 0.30), revealing distinct diagnostic approaches and accuracy levels, shown in FIG. 8D.

[0086] These analyses, especially the stark contrast in performance between the human experts and the ECE detection system shown in FIG. 1 underline the significant potential of Al to enhance medical diagnostics by providing more reliable and accurate ECE assessments.

[0087] The preoperative detection of ECE significantly influences the surgical approach during RARP, directly impacting the postoperative recovery of erectile function. ECE, an adverse pathological sign present in a third of prostate cancer patients at diagnosis is associated with increased rates of positive surgical margins and earlybiochemical recurrence post-prostatectomy. This underscores the critical need for accurate preoperative ECE prediction to tailor surgical plans and ensure safe margin resection without compromising functional outcomes. Despite the validation of clinical nomograms for ECE prediction, these were developed prior to the widespread adoption of mpMRI, highlighting a gap in leveraging modem imaging in surgical planning.

[0088] In summary, the foregoing demonstrates the value of multi-slice image analysis through a special-purpose deep learning model, e.g., first DL model 126 in FIG. 1, providing a richer dataset and accommodating the variability inherent in MRI scans. The second deep learning model 130 at the patient level revealed an AUC of 0.88 and ACC of 0.82, markedly surpassing the diagnostic accuracy of radiologists and traditional radiology reports, as shown in FIGs. 8A-8D. This indicates the model's proficiency in detecting nuanced cues within MRI slices — cues that may elude even expertly trained eyes. Incorporating texture analysis-based radiomics or the like could further refine the ability to distinguish ECE, supporting the potential of a multifaceted Al approach in augmenting diagnostic precision. In some embodiments, when implementing a decision tree model, clinical parameters such as the Gleason score and age are integrated to highlight the synergistic potential of melding clinical factors with Al predictions for detecting ECE.

[0089] In some embodiments, a patient pool can be derived from one single clinical practice, despite the MRI images being obtained from several different machines. The images are classified according to the staging information obtained from the final surgical pathology without reanalyzing the slides. External validation remains a future priority to ensure the model's effectiveness in clinical practice, and expanding the datasetto include a broader range of MRI sequences could further enhance its diagnostic capabilities.

[0090] Accordingly, not only does the ECE detection system in accordance with embodiments facilitate accurate ECE detection, but it also supports nuanced clinical decision-making, impacting treatment strategies and patient outcomes. Looking ahead, expanding external collaborations to validate the findings above and incorporating additional imaging sequences could bolster the model's utility across diverse clinical settings. By integrating sophisticated DL models with a streamlined image processing framework, the diagnostic workflow according to the embodiments above offer a web application that enables quick, reliable ECE identification. This initiative has the potential to improve diagnostic accuracy and consistency, promoting early intervention and precise treatment planning, thereby enhancing patient care.

[0091] While the invention has been described with reference to certain embodiments, it will be understood by those skilled in the art that various changes may be made, and equivalents may be substituted for elements thereof to adapt to particular situations without departing from the scope of the disclosure. Variations in convolutional neural network (CNN) architecture, including but not limited to changes in the number of layers, types and configurations of convolutional filters, pooling strategies, activation functions (e.g., ReLU, LeakyReLU, GELU), normalization techniques (e.g., batch normalization, layer normalization), or dropout rates may also be contemplated within the scope of the invention. Therefore, it is intended that the claims are not limited to the particular embodiments disclosed, but that the claims will include all embodiments falling within the scope and spirit of the appended claims.

Claims

What is claimed is:

1. An image processing system for diagnosing a medical condition, comprising: an input for receiving a plurality of images of an anatomical region from a medical instrument; a preprocessing system that standardizes the images to have common dimensions; a first deep learning model identifying and selecting diagnostically significant image slices of the standardized images; and a second deep learning model that evaluates the diagnostically significant image slices for abnormalities indicative of disease at the anatomical region.

2. The system of claim 1, wherein the medical condition includes prostate cancer, the anatomical region is a prostate gland, and the image processing system is constructed and arranged to diagnose a presence of extracapsular extension (ECE) at the prostate gland.

3. The system of claim 1, wherein the second deep learning model classifies the diagnostically significant image slices as ECE positive (ECE+) images or ECE negative (ECE-) images.

4. The system of claim 1, wherein the first and second deep learning models each includes a convolutional neural network (CNN) algorithm implemented via executable software stored on non-transitory machine-readable media.

5. The system of claim 4, wherein the CNN algorithm is configured to include three primary depthwise separable convolution blocks, each followed by batch normalization and ReLU activation, with subsequent downsampling via max pooling, wherein the first block begins with a convolutional layer configured to retain the number of channels, followed by a pointwise convolution that expands the number of channels, wherein the pattern of channel management and spatial reduction is mirrored in subsequent blocks with channels doubling from the output of the previous block, wherein the max pooling reduces spatialresolution to reduce the input's dimensionality, wherein after the final pooling layer, the CNN algorithm transitions to a dense segment consisting of two fully connected layers separated by a dropout layer set at 60% to mitigate overfitting, and a final layer outputting a probability distribution used by the first or second deep learning model for classifying an output.

6. The system of claim 1, wherein the first deep learning is trained by an output of the preprocessing system.

7. The system of claim 1, wherein the images are MRI images, computed tomography (CT) scans, Positron Emission Tomography (PET) scans, X-rays, or ultrasound images.

8. The system of claim 1, wherein the plurality of images include two dimensional (2D) image slices, and wherein the preprocessing system comprises a converter that converts the 2D image slices from a Digital Imaging and Communications in Medicine (DICOM) format to a Portable Network Graphics (PNG) format.

9. The system of claim 8, wherein each 2D image slice is taken from an axial, coronal, or sagittal plane, and when combined, the slices provide a three- dimensional representation of the scanned area.

10. The system of claim 1, wherein the preprocessing system comprises: a padding module that standardizes the images to have square dimensions; and a harmonizing module that standardizes image characteristics including intensity, contrast, and resolution across the images.

11. The system of claim 1 , further comprising a cropping module between the first deep learning model and the second deep learning model for cropping thediagnostically significant image slices so that a region of interest (ROT) in the anatomical region arc prevalent for the second deep learning model.

12. The system of claim 11, wherein an output of the cropper is used for training the second deep learning model.

13. The system of claim 1, wherein the images are input as pre-sliced two dimension inputs or three dimension inputs, wherein the system further comprises a script that performs a three dimension to two dimension slicing operation.

14. An image processing system for diagnosing a possible preoperative staging of prostate cancer (PCa), comprising: an input for receiving a plurality of images of a prostate gland from a medical instrument; a preprocessing system that standardizes the images to have common dimensions; a first deep learning model identifying and selecting diagnostically significant image slices of the standardized images; and a second deep learning model that evaluates the diagnostically significant image slices for a presence of extracapsular extension (ECE) at the prostate gland.

15. The system of claim 14, wherein the preprocessing system comprises: a padding module that standardizes the images to have square dimensions; and a harmonizing module that standardizes image characteristics including intensity, contrast, and resolution across the images.

16. The system of claim 14, wherein the first and second deep learning models each includes a convolutional neural network (CNN) algorithm implemented via executable software stored on non-transitory machine-readable media.

17. A method comprising : uploading an image file, wherein the image file is a three-dimensional magnetic resonance image (MRI) that includes a prostate image of a prostate; multi-slicing the image along one or more anatomical orientations to provide a plurality of two-dimensional image slices; harmonizing each sliced image to generate a harmonized image with standardized dimensions, wherein the harmonization includes padding the two- dimensional images using a script; utilizing a first deep learning model with a first convolutional neural network (CNN) architecture to group the harmonized images into (1) distinct slice groups that display the prostate and (2) non-distinct slice groups that do not; cropping each harmonized image in the distinct slice group around a region of interest (RO I) that includes the prostate, thereby producing cropped images; and employing a second deep learning model that is a second CNN trained to classify the cropped images as either extracapsular extension positive (ECE+) or extracapsular extension negative (ECE-).

18. The method as recited in claim 17, further comprising quantifying a percent probability of ECE+ for each of the cropped images.

19. The method as recited in claim 17, further comprising quantifying an aggregate ECE+.

Images