Systems and methods for mammalian transfer learning

Abstract

Neural networks are trained using transfer learning to analyze medical image data, including 2D, 3D, and 4D images and models. If the target medical image data is associated with a species or problem class for which there is not enough labeled data available for training, the system can create an enhanced training dataset by selecting labeled data from other species and / or different problem classes. During training and analysis, the image data is divided into portions large enough to preserve meaningful context for the problem class, yet small enough to obscure the species source (e.g., image portions small enough that it is impossible to determine whether they are from a human or a dog, but abnormal liver tissue can still be identified). The trained checkpoints can then be used to provide automated analysis and heat mapping of input images via a cloud platform or other application.

Description

[Technical field]

[0001] This application is a non-provisional extension of U.S. Provisional Application No. 62 / 987,441, filed March 10, 2020, and entitled “Systems and Methods of Mammalian Transfer Learning,” the entire disclosure of which is hereby incorporated by reference herein. [Background technology]

[0002] Machine learning problems involving 3D image classification or segmentation often utilize large amounts of high-quality, diverse labeled data to enable the creation of neural networks that have a canonical statistical sample of the problem space and can accurately perform a given task. This problem is exacerbated in parts of the medical imaging space, for both human and veterinary applications, due to the unique challenge of providing expert labeled data for complex problems. Some data may be available in a canonical labeled format primarily for human applications (e.g., for radiology), while for other problems (e.g., histopathology), more comprehensive datasets exist for animals (e.g., pigs in the case of histopathology). Improved systems and methods for transfer learning are needed to broaden the applicability of machine learning techniques to human or veterinary applications that lack sufficient labeled data.

[0003] The present disclosure will be more readily understood from the following detailed description of several illustrative embodiments taken in conjunction with the figures, in which: [Brief description of the drawings]

[0004] [Figure 1] FIG. 1 is a schematic diagram of an exemplary system configured for mammalian transfer learning. [Diagram 2] 1 is a flowchart of an exemplary set of steps that may be performed during a first session of a mammalian transfer learning process. [Diagram 3]1 is a flowchart of an exemplary set of steps that may be performed during an optional second session of a mammalian transfer learning process. [Figure 4] FIG. 2 is a schematic diagram showing possible combinations of training data. [Diagram 5] 1 is a flowchart of an exemplary set of steps that may be performed during a training and validation process. [Figure 6] 1 is a flow chart of an exemplary set of steps that may be performed during testing and application. [Figure 7] 1 is a screen shot of an exemplary interface for viewing medical image data including heat map identification of target anomalies. [Figure 8] 13 is a screen shot of another exemplary interface for viewing medical image data including heat map identification of target anomalies. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0005] Various non-limiting embodiments of the present disclosure will now be described to provide an overall understanding of the principles of the structure, function, and use of the systems and methods as disclosed herein. One or more examples of these non-limiting embodiments are illustrated in the accompanying drawings. Those skilled in the art will appreciate that the systems and methods specifically described herein and illustrated in the accompanying drawings are non-limiting embodiments. Features shown or described with one non-limiting embodiment may be combined with features of other non-limiting embodiments. Such modifications and variations are intended to be included within the scope of the present disclosure.

[0006] References throughout this specification to "various embodiments," "several embodiments," "one embodiment," "several example embodiments," "one example embodiment," or "an embodiment" mean that a particular feature, structure, or characteristic described in connection with any embodiment is included in at least one embodiment. Thus, appearances of the phrases "in various embodiments," "in some embodiments," "in one embodiment," "several example embodiments," "one example embodiment," or "in an embodiment" in various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0007] Throughout this disclosure, references to components or modules generally refer to items that may be logically grouped together to perform a function or a group of related functions. Like reference numbers are intended to generally refer to the same or similar components. The components and modules may be implemented in software, hardware, or a combination of software and hardware. The term software is used in various ways to include not only executable code, but also data structures, data stores, and computing instructions in any electronic format, firmware, and embedded software. The terms information and data are used in various ways to include a wide variety of electronic information, including but not limited to machine-executable or machine-interpretable instructions, content such as text, video data, and audio data, among others, and various codes or flags. The terms information, data, and content are sometimes used interchangeably when the context permits.

[0008] The examples discussed herein are illustrative only and are provided to aid in the description of the systems and methods described herein. None of the features or components shown in the drawings or discussed below should be understood as essential to any particular implementation of any of these systems and methods, unless specifically named as essential. For ease of reading and clarity, certain components, modules, or methods may simply be described with certain figures. Failure to specifically describe any combination or subcombination of components should not be understood to indicate that any combination or subcombination is not possible. Also, with respect to any method described, whether the method is described in conjunction with a flow diagram or not, unless otherwise specified or required by the context, it should be understood that any explicit or implicit order of steps performed during execution of the method does not imply that these steps must be performed in the order presented, but may instead be performed in a different order or in parallel.

[0009] As has been described, there is a need for a system and method for mammalian transfer learning that combines different datasets, where several quantities of data originate from different animals or problems, to enable training of a neural network that can perform the task with a high degree of accuracy. The availability of comprehensively labeled task-specific datasets is what is traditionally necessary to be able to train an artificial intelligence (e.g., an artificial neural network) to recognize and detect specified features within these datasets or new instances of the same type of data (e.g., to display the difference between healthy tissue and cancerous or abnormal growths in an MRI scan of a living organism). In producing a "high quality" training input, multiple factors are evaluated such that an AI trained using said input can output a solution to highlight anomalies along with the specificity and accuracy of the AI's assessment of the desired anomalies as specified by location labels in the medical scan. The benefit of AI is that it can operate at faster than human rates while dynamically detecting anomalies both below the range of perception of a human interpreter and without limiting factors common to humans, such as fatigue, oversaturated retinas, or insufficient knowledge of cases that humans have never encountered. Yet, with traditional approaches, the use of AI comes at the cost of improved performance, as large amounts of data are required to effectively train the AI ​​system: the size of the required datasets typically runs into the tens or hundreds of thousands, if not millions, of data points.

[0010] Labeled data can refer to data that is added to a primary image or information (e.g., CT scan, MRI, x-ray, ultrasound, fMRI, multi-modality 2D, multi-modality 3D, multi-modality 4D) and stored in a separate file from the primary information. Labeled data can be stored in a digital file format, such as, but not limited to, .dicom, .dcm, .nii, nifti, .mhd, .h, .jpg, .mpeg, or any combination herein of any multi-dimensional data, including, but not limited to, 3D or 2D data. The label can have a correspondence to the format of the primary data and a spatial orientation with the value contained within the pixel / voxel of the label. The label format can undergo file type conversion and the data can be redefined from the original value type to another value type or voxel defined class (e.g., class 0=background, class 1=organ, class 2=cancer). Voxel-defined class values ​​are not limited to a set number of units or classification types in the label file, but can be dynamically assigned, sorted, and disseminated to any number of data classifications named by the user. As one illustrative example, a label may be a set of metadata that corresponds to a set of image data, where such correspondence is expressed as a pixel-by-pixel or voxel-by-voxel association of some property not explicitly present in the image data.

[0011] The labeled data can come from multiple machines, methods, and subjects. Label attributes can include, but are not limited to, species, label type, and label class options or values. Each data comes from a single subject species. Label types can include, but are not limited to, classification, bounding box regression, and segmentation. Classification label values ​​are categorical values, bounding box regression label values ​​are coordinate centers and box dimensions that represent the location of objects in the input data, and segmentation is, for example, a binary heatmap or heatmap with categorical class values ​​for each voxel. Labeled inputs within this technique will focus on subjects of the data that will be referred to as "humans," "animals," and "alternate animals," where alternative animals refer to different animal species (e.g., dogs, pigs, horses, cats, and rodents) that are different from the previous ones that will be input to the training data, and "species" can refer to any single animal in the set of mammals. Label class options and values ​​can include, but are not limited to, specific tissues, organs, organelles, and anomalies.

[0012] Comprehensive labeled human medical datasets (including but not limited to modalities such as CT, MRI, X-ray, ultrasound, and histopathology) are available in greater quantities than labeled animal-based veterinary data of similar modalities, due to the relative scarcity of labeled animal data. A relative scarcity of sufficient labeled data also exists when translating abundant animal-derived data to human-derived data, depending on the problem or modality considered. For example, there is quantitatively more labeled histopathology data for animal subjects (mainly pigs) than for humans. To address the discrepancy in the quantity of high-quality labeled data within the medical and veterinary image analysis problem space according to the present disclosure, labeled data from a particular species can be used to train an AI to detect anomalies, supplemented fully or partially from an alternative species (e.g., adding labeled human data to help detect anomalies in dogs). To facilitate this supplementation, modality and category labels can be preserved throughout all selected sets in all training, validation, and testing sessions.

[0013] Some implementations of the disclosed technology may include, for example, a cloud-based platform that may be accessed by medical professionals, patients, insurance companies, or other parties seeking analysis of radiological scans. Although FIG. 1 illustrates a schematic of an artificial intelligence-based case study analysis computing system 10 that may be accessed by a user device 12 via a network 14 (e.g., a LAN, a WAN, or other communications network), it should be understood that the disclosed technology may also be implemented as a standalone locally executed software application. The computing system 10 may include one or more servers (e.g., physical servers, virtual servers, cloud servers, or other computing environments), each of which may include a processor, memory, storage devices, communications devices, graphical processors, and other components as are commonly used in data transfer, manipulation, and analysis. The user device 12 may be, for example, a computer, a laptop, a smartphone, a tablet, a hospital information system, a medical imaging device (e.g., an MRI machine or other digital imaging equipment), or other device capable of communicating with the computing system 10.

[0014] An artificial intelligence-based case study analytical computing system can be trained on labeled mammalian data, such as human-only data or a combination of human and animal data. In such an instance, the remote user (12) may be, for example, a veterinary information system seeking analysis of scans of animal organs to aid in diagnosis.

[0015] As shown, a remote user (12) can begin the process with examining a patient and collecting a digital diagnosis, such as one or more scans of one or more organs. The user can form a case study of the animal under examination and upload the case study to an artificial intelligence based case study analysis computing system. The content of the uploaded case study can vary, but the case study can include one or more scans of a particular organ or organs of the animal, as collected by veterinary imaging equipment. Such uploading can occur through any suitable data transmission technique, such as, for example, through an online web-based portal, a software application interface, or other communication channel.

[0016] Upon receiving the case study from the remote user, the artificial intelligence based case study analysis computing system can perform a case study analysis based on its imaging model and add the analysis to the case study. The analysis of the process can be added to the case study using any of a number of suitable approaches. In one embodiment, for example, the artificial intelligence based case study analysis computing system zooms in on the case study with a color-coded heat map to visually indicate potential problem areas. The case study with the analysis can then be downloaded or otherwise transmitted or provided to the remote user. The remote user can take appropriate next steps based on the analysis of the case study provided by the artificial intelligence based case study analysis computing system.

[0017] According to various embodiments, the systems and methods can be provided as a software platform or application that can analyze past case studies to form a reference point for abnormal findings and report these findings. Thus, a more diagnostically complete visual analysis of a case can be provided without limiting the ultimate findings to a particular disease that is visually present within predefined intrinsic parameters.

[0018] Labeled data enters the system through the process depicted in FIG. 2, which illustrates generally an example process flow for both training, validation, and testing sessions for transfer learning according to the present disclosure. As illustrated, the process flow according to the present disclosure can have various levels of tuning and testing, which may depend on the desired level of accuracy sought, resource availability, and / or other operating parameters. According to various embodiments, the systems and methods described herein using one or more of the process flows depicted in FIGS. 2-3 as non-limiting use case examples can achieve a confidence level of at least 70% in some embodiments, while other embodiments can achieve a peak confidence level of greater than 95%. In additional non-limiting use case examples, the level of accuracy in cases of liver, lung, kidney, and spleen masses / tumors in 20-80 lb. cats and dogs can exceed 80%, while in other embodiments the level of accuracy can exceed 85% or 87%.

[0019] Referring to FIG. 2, the single-session method (100) controls when the single-session method is applied to the first training session. The single-session method allows the user to train using data outside the scope of the species application list, which refers to the intended use case of the model in the field as long as the same class labels are used within the same instance. The single-session method refers to object detection methods and instance segmentation models that undergo a configurable window or type of rearrangement of values ​​while sampling for region proposals. The model gains the ability to stratify labeled input data from multiple sources before applying the target data by splitting the labeled input data into "chunks" defined as subsections of the original scan / information that are small enough to obscure the species of the data instances while still remaining large enough for the model to classify internal organs, anomalies, and other feature classes. These subsets will be referred to as "chunks" throughout the remainder of this document.

[0020] In Figure 2, the system may select mixed source data for the session (102a) for which the method is configured for a single session (100). A mixed source dataset means that data from two or more species and one or more classes are used to form a dataset that can later be split into training, validation, and test sets. The user should also specify the application list of species that will later be used by the advanced sampling method. Mixed source data can originate from, but is not limited to, scenarios in which labeled human training data is applied to animal test data, labeled animal training data is applied to animal test data, labeled animal training data is applied to alternate animal test data, labeled animal training data is applied to human test data, a combination of labeled human and animal training data is applied to animal test data, a combination of labeled human and animal training data is applied to human test data, a combination of labeled human and animal training data is applied to alternate animal test data, a combination of labeled human, animal and alternate animal training data is applied to animal test data, a combination of labeled human, animal and alternate animal training data is applied to human test data, or any combination of the foregoing, and the "alternate animal" data can refer to one or more animal species that are not the same species as the first animal (e.g. animal = dog, alternate animal = cat and pig).

[0021] The distribution of source data and corresponding applications is shown in Figure 4, which shows how different data sources are collated and used together in the transfer learning process. This figure shows that different applications can be combined together to form an integrated solution. Applications can be defined as different data modalities (including but not limited to T1 or T2 scans in the MRI field), scan types (including but not limited to ultrasound, MRI, or CT scans), labels or object types, or different formats of data that can be shared to a system including but not limited to problems arising from human or veterinary solutions (including but not limited to different medical problems such as cancer or other medical abnormalities), where an animal application can consist of modalities or scans arising from different species of animals.

[0022] In FIG. 2, the system is configured for two or more single sessions (100), and selecting source-specific data for a session (102b) is where the user selects a single species and one or more classes to form a data set that can later be separated into training, validation, and test sets. The species application matches the species selected by the user. Some examples of source data can come from scenarios such as, but not limited to, labeled human training data is applied to human test data, labeled animal training data is applied to animal test data, and labeled surrogate animal training data is applied to surrogate animal test data.

[0023] Once the mixed domain data is selected (102a), the system may select an advanced sampling method (104a), which may control which species are sampled for different data sets using a stratified sampling method. The user should provide parameters to control the sampling method. The configuration should be selected to allow any of the source data to be used for training, while only allowing target data that match the target application list to be used for test and validation sets.

[0024] Once the domain-specific data is selected (102b), the system may select a conventional sampling method (104b), which may include, but is not limited to, a simple random sampling method and a systematic sampling method; if systematic sampling is selected, the user should also provide parameters to control the sampling method.

[0025] The objective of the pre-processing methods shown in this disclosure in Figures 2 and 3 is to minimize the volume and / or channels of the input image while maximizing the capacity of the input features. Pre-processing helps differentiate the examples into their class label groups when processed with the trained learning model. In this way, the system may select a pre-processing method (106), which may include, but is not limited to, feature normalization, image scaling, feature engineering, and 2D to 4D custom versions of sliding window methods, which will be referred to as sliding window methods. Sliding window methods are used for classification and semantic segmentation models, such as but not limited to u-net, which is used to demonstrate the transfer learning methods described herein to simulate the impact of region proposal methods or similar alternatives. This restructuring to a part-to-whole window method makes the data between different species more and more similar. The smaller the data subsets, windows, or "chunks," the greater the similarity between the labeled input data and the target data. Still, windows are only useful if the window is large enough for the model to correctly differentiate class labels using the windowed input images. Other pre-processing methods change the scale of the image before the window is applied, resulting in changes to the optimal window size. Different methods produce different types of inputs, which affects all possible processing of information by the AI. Post-processing is necessary when sliding window methods are used on models to reconstruct the final output. In the case of a classification network, this should result in the output being condensed into a single output class vector.

[0026] Model selection involves choosing a learning model for implementing any version of transfer learning. Transfer learning can be applied to all supervised learning models and model types for image processing. Model selection (108) is the process of choosing one of these model types. Examples can include, but are not limited to, classification, semantic segmentation, object detection, and instance segmentation. Different model types can require data with different groups of values ​​of label types. A classification model data example requires a classification label type. A semantic segmentation model data example requires a segmentation label type. An object detection model data example requires classification and bounding box regression label types. An instance segmentation model data example requires classification, bounding box regression, and segmentation label types. Examples of classification models include, but are not limited to, VGG, ResNet, SE-ResNet, SENet, DenseNet, Inception Net, Mobile Net, EfficientNet, and Alex Net. Sub-networks can be added to classify data with a time dimension. Examples include, but are not limited to, RNN, LSTM, or GRU. Examples of semantic segmentation models include, but are not limited to, u-net, Linknet, PSPNet, and FPN. Examples of object detection models include, but are not limited to, r-cnn, fast r-cnn, faster r-cnn, RetinaNet, and YOLO. Examples of instance segmentation models include, but are not limited to, YOLACAT, mask r-cnn, and DETR. Object detection and instance segmentation model types decompose the image into smaller regions before computing classes, bounding box updates, and / or segmentation masks. When diagnostic images of humans and other mammalian species are diced into smaller images, it is difficult to distinguish which chunks originate from which species. The same phenomenon can arise from classification or semantic segmentation models when sliding window algorithms are implemented in the pre-processing selection.The implementation of the disclosed model that will be addressed in the context of this document is a u-net that performs semantic segmentation using a custom implementation of a sliding window in the pre-processor.

[0027] User selection of settings (110) may include, but are not limited to, data split ratios (training, validation, and test sets sampled from selected domains and classes), loss function, optimizer, learning method, batch size if applicable, and number of epochs, as shown in FIG. 2. Settings may also optionally include data augmentation methods, baselines, and hyperparameters. While these settings are not directly related to the unique application and method of transfer learning described herein, the selection of these settings is related to the effectiveness of any model to fit data, and such selection will be apparent to one of ordinary skill in the art in view of this disclosure.

[0028] Regarding the data split ratio in the setup, another point of division should be that the majority of the data is utilized for this application as test / validation versus training dataset size. This includes the number of supplemental data cases selected for the target case being solved. For example, if the target case is feline liver cancer, the supplemental data cases selected for training include, but are not limited to, human liver cancer, canine liver cancer, and human adrenal mass cancer. Non-limiting examples of ratios of target cases to supplemental cases include 2 other cases for 1 target case, 3 other cases for 1 target case, 4 other cases for 1 target case, 5 other cases for 1 target case, 6 other cases for 1 target case, 7 other cases for 1 target case, 8 other cases for 1 target case, 9 other cases for 1 target case, 10 other cases for 1 target case, 11 other cases for 1 target case, 12 other cases for 1 target case, 13 other cases for 1 target case, 14 other cases for 1 target case, 15 other cases for 1 target case, 16 other cases for 1 target case, 17 other cases for 1 target case, 18 other cases for 1 target case, 19 other cases for 1 target case, and 20 other cases for 1 target case.

[0029] Examples of data augmentation methods may include, but are not limited to, random reordering of slices about one of the x, y, or z axes, some intensity blurring applied to the slices, manipulation of the scale of one of the x, y, or z axes. Data augmentation methods are only applied to the training data set of a session as applied in a typical use of the technology.

[0030] Examples of benchmarks include, but are not limited to, fl score, precision, recall, specificity, accuracy, DICE coefficient, and other confusion matrix metrics and their multi-class confusion matrix equivalents.

[0031] Examples of loss function settings include, but are not limited to, mean squared error, mean absolute error, hinge loss, and cross entropy loss.

[0032] The optimizer may use the training data to generate iterative updates to the model to reduce the error score generated by the loss function. Examples of optimizers include, but are not limited to, Momentum, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, Adam, AdaMax, Nadam, and AMSGrad.

[0033] The training method controls how much of the dataset is considered to compute the model updates. Examples of training methods include, but are not limited to, stochastic gradient descent, batch gradient descent, and mini-batch gradient descent.

[0034] If the user selects batch gradient descent or mini-batch gradient descent for the training method in the setting, the user will need to supply a batch size. Examples of batch sizes may include, but are not limited to, 1, 2, 5, 10, 12, 15, 20, 24, 30, 32, 36, 40, 42, 45, 50, 52, 60, 64, 70, 72, 80, 82, 84, 90, 96, and 100.

[0035] Epochs refer to the number of times a session iterates through the training and validation data sets. Implementations of the disclosed system can arrive at a meaningful solution by training the neural network using various numbers of epochs, with various implementations including about 3 epochs, about 5 epochs, about 7 epochs, about 10 epochs, about 15 epochs, about 20 epochs, about 25 epochs, about 30 epochs, about 35 epochs, about 40 epochs, about 45 epochs, about 50 epochs, about 60 epochs, about 70 epochs, about 80 epochs, about 90 epochs, and so forth. Other methods include, but are not limited to, about 100 epochs, about 150 epochs, about 200 epochs, about 250 epochs, about 300 epochs, about 350 epochs, about 400 epochs, about 450 epochs, about 500 epochs, about 550 epochs, about 600 epochs, about 650 epochs, about 700 epochs, about 750 epochs, about 800 epochs, about 850 epochs, about 900 epochs, about 950 epochs, and about 1,000 epochs. Other methods allow the number of epochs to be determined during a session by setting a condition on how comparable the training and validation baseline histogram scores are.

[0036] Hyperparameters control how the model or its settings can be modified iteratively during validation to search for a better model for the problem. Examples can include, but are not limited to, alternative settings, model properties such as the activation function used, the number of deep and wide layers, layer kernels and kernel sizes, learning rates, dropout, regularization, and any auxiliary output layers.

[0037] In Figure 2, running session (112) runs training, validation, and test sessions using the selected data, model, and settings. This should produce histogram values ​​of the loss function and metrics for both the training and validation steps, as well as error and metrics scores for the test step. The training and validation cycles will update the model for a specified number of epochs in some embodiments or in other embodiments, until user-specified goals in loss and metrics for the training and validation histograms are met.

[0038] The dual session decision block is where the user can choose, at his / her discretion, to end the process flow after the first session or to continue with a second session (114). If the session is a single session method (114), the process ends and the application is trained and ready to be used to analyze case studies from real patients. If the session is a dual session method (114), the system can proceed to the steps of Figure 3. The single session decision block (200) of Figure 3 is where the user can choose to use the dual session method by itself or in conjunction with the single session method from Figure 2 for the second session, regardless of whether the single session method was previously used in the first session.

[0039] For the dual session method shown in Figure 3, the system can select mixed target data for the session (202a), indicating that the first session method is used for the second session. A mixed target dataset means that data from two or more species and one or more classes are used to form a dataset that can later be separated into training, validation, and test sets. The user should also specify the application list of species that will later be used by this or more classes to form a dataset that can later be separated into training, validation, and test sets. The mixed target data can originate from, but is not limited to, scenarios where labeled human training data is applied to animal test data, labeled animal training data is applied to animal test data, labeled animal training data is applied to alternative animal test data, labeled animal training data is applied to human test data, a combination of labeled human and animal training data is applied to animal test data, a combination of labeled human and animal training data is applied to human test data, a combination of labeled human and animal training data is applied to alternative animal test data, a combination of labeled human, animal and alternative animal training data is applied to animal test data, a combination of labeled human, animal and alternative animal training data is applied to human test data, or any combination of the foregoing for the first session.

[0040] For the dual session method shown in Figure 3, the system can select target specific data for a session (202b), which can include a case where a user selects a single species and one or more classes to form a data set that can later be separated into training, validation, and test sets. Some examples of target data can come from, but are not limited to, scenarios where labeled human training data is applied to animal test data, labeled animal training data is applied to animal test data, labeled animal training data is applied to alternate animal test data, labeled animal training data is applied to human test data, a combination of labeled human and animal training data is applied to animal test data, a combination of labeled human and animal training data is applied to human test data, a combination of labeled human and animal training data is applied to alternate animal test data, a combination of labeled human, animal, and alternate animal training data is applied to animal test data, a combination of labeled human, animal, and alternate animal training data is applied to human test data, or any combination herein.

[0041] If mixed-domain target data is selected (202a), the user may select an advanced sampling method (204a), which may include the user selecting two configurations of a stratified sampling method to control which species are sampled for different data sets. The configuration should allow any of the target data to be used for training while only allowing data that matches the target application list.

[0042] If domain-specific target data is selected (202b), the user may select a conventional sampling method (204b), which may include, but is not limited to, simple random sampling and systematic sampling methods.

[0043] In either case, the system may prepare the pre-trained model for the dual-session mammalian transfer learning method (206), which involves two optional steps: selecting an early model layer to freeze for the next learning session, and selecting a later model layer to replace the randomly initialized layer. Such selection may be performed manually, but the present disclosure is not limited to manual selection alone. While replacing the final layer of the model is typically required for traditional transfer learning, generalizable results can be achieved without this step, since the output classes of the disclosed application are preserved between sessions. In general, the more layers frozen in the model, the less target species data is required to fit the model to the problem space as a result of using the dual-session method described herein. As with the single-session method, the dual-session method may require that class categories remain constant while species may vary.

[0044] The Modify Settings process (208) from Figure 3 is the point where the user can change the settings previously used in the first learning session in order to optimize the generalizable scores or outputs from the previous session. The hyperparameter selection at this point is limited to changes that do not change the pre-trained model architecture or pre-processing methods, and therefore should not make the resulting system immovable or with meaningless outputs.

[0045] In Figure 3, running session (210) performs training, validation, and testing sessions using the selected data, model, and settings. This process should produce histograms of the loss function and baselines for both the training and validation cycles, and then error and baseline scores for the testing step. The training and validation cycles will update the model using the loss function, optimizer, and other required settings for the specified number of epochs (or until user-specified goals of the loss and baselines for the training and validation histograms are met).

[0046] FIG. 5 illustrates a set of steps that may be performed during training and validation, such as may occur in connection with FIGS. 2 and 3 (e.g., as part of a session being run or executed (112, 210)). Although FIGS. 5 and 6 illustrate certain steps being performed by a "CPU" and a "GPU," it should be understood that this is merely illustrative and that various steps may be performed by any processing unit or by a combination of CPU, GPU, TPU, and other processing units.

[0047] The steps of FIG. 5 begin with providing and / or receiving one or more labeled mammal datasets (300). The system selects (302) and blends one or more labeled mammal datasets and stores (304) the selected data. The stored data is pre-processed (306) and converted to a standard format, and the ready data is stored (308). Parameter selection is configured (310) for the network and data, and parameters are loaded for use (312). The system then performs data segmentation, splitting, and blending of the ready data (314) to define segments, portions, or chunks of the ready data, which are then stored (316). Defining these chunks may include the creation of a new file or dataset for each chunk, or may include the creation of metadata that can be applied to the ready data to identify each chunk. The system then trains (318) the neural network using the chunks of data and stores (320) the model parameters, resulting baselines, and loss characteristics. The system may utilize an optimizer to fine-tune the performance of the neural network based on the stored (320) characteristics and may determine and store (324) one or more improved parameters. The system may perform (326) validation of the training with the neural network to determine (328) an independent set of results.

[0048] The system may then determine (330) whether a sufficient number of epochs have been performed and whether a sufficient level of accuracy has been achieved during training of the neural network by comparing (328) the independent results to standard results of neural networks with similar input data sets. If the accuracy is insufficient, the system may proceed to subsequent epochs by training (318) the neural network on chunks with improved parameters (324). If the accuracy is sufficient (330), the system may produce a checkpoint (332) of the neural network that is available for testing and independent analysis of the input data set.

[0049] FIG. 6 illustrates a set of steps that may be performed during testing of checkpoint versions and use of a neural network to analyze real case studies. The system may access or receive one or more mammalian datasets (400). These may be test datasets that can be used to validate the validity of the checkpoint versions or may be real-world case studies received from a user of the system. The system may select a checkpoint to process data (402) and may access a checkpoint where the data is stored (404). The system may perform data selection (406) to select cases that share some commonality with the target data and may load the test data (408). Network and data parameters may be configured and / or selected (410) and these configurations may be loaded into the neural network (412). The system may then perform data segmentation (414) to separate the target cases and related cases (416) into subsets. As with the prior example, this may involve creating a new dataset for the chunks, or creating a set of metadata that defines the chunks within these original datasets. The system then applies the checkpoint neural network to the input cases (418) and produces a heatmap output (420).

[0050] The heatmap output may include pre-rendered graphics or models, or may include metadata that can be used to overlay or insert visual indicators into the rendered image. Figures 7 and 8 each provide examples of 3D models with identified target tissues marked. Figure 7 also visually shows the chunks into which the image has been divided, while other rendered heatmaps, such as that of Figure 8, may show the individual chunks. The system may then analyze (422) the heatmap output to generate a set of accuracy metrics (424) and provide a completed result (426) to the user.

[0051] According to the present disclosure, labeled mammal data of one or more species is utilized during training of an AI network model to detect anomalies of mammalian species. In some embodiments, machine learning network training is performed using exclusively labeled human data. In other embodiments, machine learning network training is performed using a combination of labeled human data and labeled animal (i.e., non-human) data. According to various embodiments, when the network is tested, the labeled animal data is used to ensure that the reported accuracy of the network is focused on how well it generalizes to the problem space of the animal target. In any event, the systems and methods of the present disclosure can leverage the similarity of the labeled mammal data to help compensate for the lack of data necessary to properly learn to select and distinguish abstract features, which allows for a generalized network model solution.

[0052] The methods disclosed herein advantageously enable rapid development of machine learning solutions, including but not limited to human radiology, among other imaging modalities, using learning developed from animal-based datasets or other high-quality dataset sources. Thus, the systems and methods described herein can be utilized across a variety of disciplines, including but not limited to general practice, radiologists, medical specialists, diagnosticians, and imaging specialists, collectively referred to herein as "healthcare professionals."

[0053] Once the network has been trained using the process outlined in the previous section, the network trained using transfer learning can be applied to live data in the field. In one embodiment, data would be received in the form of a 2D or 3D scan. The data in this embodiment would be separated into chunks containing portions of the scan and fed through the network. The network would output a confidence value for each voxel reconstructed into a heatmap matching pattern of the original scan. This heatmap would be refined by the use of intervals to select all or a portion that would be designated as an identified anomaly. This portion of the heatmap would then be overlaid on the original scan to provide an assessment of where the medical problem is.

[0054] According to various embodiments, the systems and methods can be used to help optimize the model when further provided with appropriate pre-processing, sampling methods, settings, and data sets. As one illustrative example, this can include a cloud platform or local software application that analyzes past case studies to form a reference point for abnormal findings, and can report on these findings, no matter how small. Thus, a more diagnostically complete visual analysis of the case can be provided without limiting the ultimate findings to specific diseases that are visually present within the given intrinsic parameters.

[0055] In general, it will be apparent to one skilled in the art that at least some of the embodiments described herein can be implemented in many different embodiments of software, firmware, and / or hardware. The software and firmware code can be executed by a processor or any other similar computing device. The software code or specialized control hardware that can be used to implement the embodiments is not limiting. For example, the embodiments described herein can be implemented in computer software using any suitable computer software language type, for example using conventional or object-oriented techniques. Such software can be stored on one or more suitable computer-readable media of any type, such as magnetic or optical storage media. The operation and behavior of the embodiments can be described without making specific reference to specific software code or specialized hardware components. The absence of such specific reference is possible, as it is clearly understood that one skilled in the art should be able to design software and control hardware to implement the embodiments based on the present description with little or no reasonable effort and without undue experimentation.

[0056] Moreover, the processes described herein can be executed by a programmable device, such as a computer or computer system and / or a processor. Software capable of causing the programmable device to execute the processes can be stored in any storage device, such as, for example, a computer system (non-volatile) memory, an optical disk, a magnetic tape, or a magnetic disk. Furthermore, at least some of the processes can be programmed when the computer system is manufactured or can be stored in various types of computer readable media.

[0057] It will also be appreciated that certain portions of the processes described herein may be implemented using instructions stored on one or more computer-readable media that direct a computer system to perform processing steps. Computer-readable media may include, for example, memory devices such as diskettes, compact discs (CDs), digital versatile discs (DVDs), optical disk drives, or hard disk drives. Computer-readable media may also include physical, virtual, permanent, temporary, semi-permanent, and / or semi-temporary memory storage.

[0058] A "computer," "computer system," "host," "server," or "processor" can be, for example and without limitation, a processor, microcomputer, minicomputer, server, mainframe, laptop, personal data assistant (PDA), wireless email device, cellular phone, pager, processor, fax machine, scanner, or any other programmable device configured to transmit and / or receive data over a network. The computer systems and computer-based devices disclosed herein can include memory for storing certain software modules used in obtaining, processing, and communicating information. It will be understood that such memory can be internal or external with respect to the operation of the disclosed embodiments.

[0059] The machine learning code can access memory locations pointed to by the CPU or GPU and iterate through the network architecture to decode and execute instructions stored in memory. These sessions are computationally expensive and can be performed using the GPU and CPU. The CPU or GPU can be instructed by the iterator to create locations to store data in whole or in part, variations of the data, or location / orientation link data. The data can be stored in memory, which can include but is not limited to RAM or external storage devices. The instructions determine how the network will operate and function. Once the GPU has accessed the instructions in memory, it uses the instructions to assemble and move the data through the neural network that is computed on the distributed cores of the GPU.

[0060] A neural network is constructed from weights, biases, and a set of rules that describe the effect these weights have on input data as it is fed through a set of intermediate memory states. These portions may be randomized, pre-set, or loaded from an existing checkpoint. A training algorithm, stored in RAM or secondary allocated memory locations and executed using either or both the CPU and GPU in various embodiments, modifies the weights and biases using results from the GPU's computation of the neural network, such that the network constructed from the modified weights and biases is mathematically optimized to perform more accurately on the data than previous iterations of the network were able to achieve. This process is repeated in different chunks, batches, and ways in various embodiments to allow for the creation of a network that can accurately perform the required task (i.e., identify anomalies in a reconstructed medical scan).

[0061] The network is evaluated on the GPU using a process similar to training, in that the data loaded into memory is used by the GPU to complete the computation of the neural network. The results of these evaluations are stored and used to determine the accuracy or behavior of the network, which can be further specified and elaborated in the context of an algorithm. Descriptions of such settings that control the process of how a network is trained, evaluated, and tested are called hyperparameters, and are generally designated by a user in some embodiments, and by an automated algorithm in other embodiments. The weights and biases that make up the network can be stored in a data file (including but not limited to .hdf5 format, or more broadly any other file format that contains an ordered matrix of numbers) for later reconstruction. The resulting data file, or any other data file of similar purpose, can also be reloaded and used to construct a neural network that behaves identically to the network used to generate the file.

[0062] In various embodiments disclosed herein, a single component may be replaced by multiple components and multiple components may be replaced by a single component to perform a given function or functions. Except where such substitution is not possible, such substitution is within the contemplation of the embodiments. A computer system may include one or more processors in communication with memory (e.g., RAM or ROM) via one or more data buses. The data buses may carry electrical signals between the processor and the memory. The processor and memory may include electrical circuits that conduct electrical current. The charge state of various components of the circuit, such as solid state transistors of the processor and / or memory circuits, may change during operation of the circuit.

[0063] Some of the figures may include flow diagrams. Although such figures may include specific logic flows, it will be understood that the logic flows merely provide example implementations of general functionality. Furthermore, the logic flows do not necessarily have to be executed in the order presented, unless otherwise indicated. Furthermore, the logic flows may be implemented by hardware elements, software elements executed by a computer, firmware elements embedded in hardware, or any combination thereof.

[0064] The foregoing description of the embodiments and examples has been presented for purposes of illustration and description. It is not intended to be exhaustive or to be limited to the forms described. Numerous variations are possible in light of the above teachings. Some of these variations have been discussed, and others will be understood by those skilled in the art. The embodiments were chosen and described to best illustrate the principles of the various embodiments as suited to the particular use envisaged. The scope is naturally not limited to the examples set forth herein, but may be employed by those skilled in the art in any number of applications and equivalent devices. Rather, it is hereby intended that the scope of the invention be defined by the claims appended hereto.

[0065] Applications of the technology within the area of ​​cardiology may include, but are not limited to, vertebral heart score, cardiac hypertrophy, arrhythmias, coronary artery obstruction, aortic obstruction, cardiac obstruction, aortic disease and Marfan syndrome, congenital heart disease, coronary artery disease, deep vein thrombosis, pulmonary embolism, mitral regurgitation, mitral valve malformation, tricuspid valve malformation, mitral valve stenosis, aortic stenosis, patent ductus arteriosus, ventricular septal defect, atrial septal defect, dilated cardiomyopathy, arrhythmic cardiomyopathy, heartworm disease, pulmonary valve stenosis, and tetralogy of Fallot.

[0066] Applications of the technique within the gastrointestinal field may include, but are not limited to, Esophageal Distension, Barrett's Esophagus, Esophageal Foreign Body, Pleural Gas, Pleural Fluid, Cranioventral Parenchymal Pattern, Caudorsal Parenchymal Pattern, Nodular or Miliary Pattern, Pulmonary Nodule, Gastric Distension, Gastric Foreign Body, Gastric Dilatation and Volvulus Small Intestinal Foreign Material Small Intestinal Plication, Two Populations of Small Intestine Colonic Foreign Material.

[0067] Applications of techniques within the field of the abdominal or renal region may include, but are not limited to, Hepatomegaly, Mid Abdominal Mass, Splenomegaly, Ascites, Retroperitoneal Fluid, Retroperitoneal Lymph Node Enlargement, Renal Mineralization, Renal Enlargement, Small Kidneys, Bladder Stones, Urethral Stones, Prostate Enlargement, Prostate Calcification.

[0068] Applications for techniques within the field of skeletal and thoracic regions may include, but are not limited to, cervical disc space narrowing spondylosis, thoracic-lumbar disc space narrowing, interstitial patterns, bronchointerstitial patterns, bronchial patterns, pulmonary masses, ruptured appendix, and invasive bone lesions.

[0069] Applications of techniques within the field of radiology may include, but are not limited to, lung cancer, colon cancer, colorectal cancer, cervical cancer, gastric cancer, bladder cancer, liver cancer, hepatic vascular cancer, adrenal cancer, kidney cancer, pancreatic cancer, thyroid cancer, breast cancer, ovarian cancer, prostate cancer, squamous cell carcinoma, basal cell carcinoma, melanoma, melanocytic nevi, actinic keratosis, benign keratosis, dermatofibroma, vascular lesions, esophageal adenocarcinoma, neuroblastoma, osteoarthritis, rib fractures, structural nerve damage, scoliosis, spinal fractures, herniated discs, pneumonia, pneumothorax, pulmonary ventilation, and COVID-19.

[0070] Applications of the technology within the field of neurology may include, but are not limited to, mild cognitive impairment, traumatic brain injury, concussion, Alzheimer's disease, dementia, Parkinson's disease, stroke lesions, multiple sclerosis, brain tumors, intracranial hemorrhage, degenerative lumbosacral stenosis, degenerative myelopathy, diffuse idiopathic osteoarthritis, and intervertebral disc disease.

[0071] Applications for techniques within the field of histopathology may include, but are not limited to, tumor proliferation, nuclear atypia scoring, signet ring cell detection, multiple myeloma, abnormal, mitotic, lymphocyte, macrophage, neutrophil, epithelium, mitochondrial fission, normal vs. benign vs. malignant white blood cell cancer analysis, metastatic tissue, immune-mediated hemolytic anemia, immune-mediated thrombocytopenia, anaplastic anemia, antibody-dependent, cytotoxicity, adenocarcinoma, atypical glandular cells, and lymphoma.

[0072] Applications of technology within the field of Ophthalmology may include, but are not limited to, Eye Disease, Optical Hemorrhage, Optical Anuerisms & Microanuerisms, Hard Exudates, Soft Exudates, Retinal Fundus Disease, Diabetic Retinopathy, Blindness, Macular Degeneration, Glaucoma, Optic Disc Displacement, Cataracts, Subconjunctival Hemorrhage, Amblyopia, Strabismus, Conjunctivitis, and Keratoconus.

[0073] Applications of the technology within the field of cardiology may include, but are not limited to, vertebral heart score, cardiac hypertrophy, arrhythmias, coronary artery obstruction, aortic obstruction, cardiac obstruction, aortic disease and Marfan syndrome, congenital heart disease, coronary artery disease, deep vein thrombosis, pulmonary embolism, mitral regurgitation, mitral valve malformation, tricuspid valve malformation, mitral valve stenosis, aortic stenosis, patent ductus arteriosus, ventricular septal defect, atrial septal defect, dilated cardiomyopathy, arrhythmic cardiomyopathy, heartworm disease, pulmonary valve stenosis, and tetralogy of Fallot.

[0074] Applications of techniques within the gastrointestinal field may include, but are not limited to, esophageal distension, Barrett's esophagus, esophageal foreign bodies, pleural gas, pleural fluid, cranioventral parenchymal type, caudal dorsal parenchymal type, nodular or miliary type, pulmonary nodules, gastric distension, gastric foreign bodies, gastric dilatation and axial small intestinal foreign bodies, small intestinal plication formation, two populations of small intestinal colonic foreign bodies.

[0075] Applications of techniques within the abdominal or renal area may include, but are not limited to, hepatomegaly, mid-abdominal masses, splenomegaly, ascites, retroperitoneal fluid, retroperitoneal lymphadenopathy, nephrocalcinosis, renal enlargement, small kidneys, bladder stones, urethral stones, prostate enlargement, prostatic calcification.

[0076] Applications for techniques within the field of skeletal and thoracic regions may include, but are not limited to, cervical disc space narrowing spondylosis, thoracic and lumbar disc space narrowing, interstitial patterns, bronchointerstitial patterns, bronchial patterns, pulmonary masses, ruptured appendix, and invasive bone lesions.

[0077] Applications of techniques within the field of radiology may include, but are not limited to, lung cancer, colon cancer, colorectal cancer, cervical cancer, gastric cancer, bladder cancer, liver cancer, hepatic vascular cancer, adrenal cancer, kidney cancer, pancreatic cancer, thyroid cancer, breast cancer, ovarian cancer, prostate cancer, squamous cell carcinoma, basal cell carcinoma, melanoma, melanocytic nevi, actinic keratosis, benign keratosis, dermatofibroma, vascular lesions, esophageal adenocarcinoma, neuroblastoma, osteoarthritis, rib fractures, structural nerve damage, scoliosis, spinal fractures, herniated discs, pneumonia, pneumothorax, pulmonary ventilation, and COVID-19.

[0078] Applications of the technology within the field of neurology may include, but are not limited to, mild cognitive impairment, traumatic brain injury, concussion, Alzheimer's disease, dementia, Parkinson's disease, stroke lesions, multiple sclerosis, brain tumors, intracranial hemorrhage, degenerative lumbosacral stenosis, degenerative myelopathy, diffuse idiopathic osteoarthritis, and intervertebral disc disease.

[0079] Applications of techniques within the field of histopathology may include, but are not limited to, tumor proliferation, nuclear atypia scoring, signet ring cell detection, multiple myeloma, abnormalities, segmentation, lymphocytes, macrophages, neutrophils, epithelium, mitochondrial fission, normal vs. benign vs. malignant white blood cell cancer analysis, metastatic tissue, immune-mediated, hemolytic anemia, immune-mediated thrombocytopenia, anaplastic anemia, antibody-dependent, cytotoxicity, adenocarcinoma, atypical glandular cells, lymphoma.

[0080] Applications of the technology within the field of ophthalmology may include, but are not limited to, eye diseases, photobleeds, photoaneurysms & microaneurysms, hard exudates, soft exudates, retinal fundus diseases, diabetic retinopathy, blindness, macular degeneration, glaucoma, optic disc displacement, cataracts, subconjunctival hemorrhage, amblyopia, strabismus, conjunctivitis, keratoconus.

Claims

1. 1. A method for training a neural network for medical image analysis using mammalian transfer learning, comprising: (a) receiving, by a processor, one or more comparison datasets, each of the one or more comparison datasets including labeled image data associated with a species; (b) generating, by the processor, a mixed domain dataset based on the one or more comparison datasets; (c) for each image of the plurality of images of the mixed domain dataset, (i) defining, by the processor, a plurality of chunks in the image, the size of each of the plurality of chunks being selected to obscure a source species of the image; (ii) adding, by the processor, the plurality of chunks of the image and any associated labels to a mixed-domain training dataset; (d) by said processor and using said neural network, (i) training the neural network to discriminate medical properties of case studies from a target species based on the mixed-domain training dataset and the associated labels, wherein the one or more species of the one or more comparison datasets include at least one species other than the target species; (ii) validating the neural network based on a validation data set selected from the mixed-domain data set; (e) generating, by the processor, checkpoints based on the neural network; A method comprising:

2. The associated label is (a) a set of per-pixel properties describing the pixels of the chunk; or (b) a set of per-voxel properties describing the voxels of the chunk; The method of claim 1 , comprising one of:

3. 2. The method of claim 1, wherein the one or more comparative datasets include a first comparative dataset associated with the target species and a second comparative dataset associated with a species other than the target species.

4. The method of claim 1 , wherein the plurality of images comprises one or more of a two-dimensional image, a three-dimensional image, and a four-dimensional image.

5. The processor, (a) one or more central processing units in communication with each other directly or via a network; and (b) one or more graphic processing units in communication with each other directly or via a network; The method of claim 1 , comprising one or more of:

6. (a) receiving, by the processor, the case study from a user device, the case study including a set of medical images associated with a patient; (b) defining, by the processor, a second plurality of chunks in the set of medical images, a size of each of the second plurality of chunks selected to obscure the patient's species; (c) analyzing the second plurality of chunks by the processor and using the checkpoints to identify the medical features within the set of medical images; (d) providing, by the processor, an indication of the medical characteristic within the set of medical images; The method of claim 1 further comprising:

7. providing the indication of the medical characteristic, (a) displaying, by the processor, at least one medical image of the set of medical images on the user device; (b) displaying, by the processor, a heat map including the indication of the medical characteristic with at least one medical image; The method of claim 6, comprising:

8. (a) each of the one or more comparison datasets comprises labeled image data associated with a species and one or more problem classes, the one or more problem classes describing one or more medical abnormalities; (b) the one or more problem classes of the one or more comparison datasets include at least one problem class that does not describe the medical characteristic; The method of claim 1.

9. 2. The method of claim 1, further comprising testing the checkpoint based on a test dataset selected from the mixed domain dataset, the test dataset and the validation dataset each associated with the target species.

10. The method of claim 1 , further comprising training the neural network to identify the medical feature in less than about 100 epochs.

11. 2. The method of claim 1, wherein the one or more comparative datasets comprise a first comparative dataset associated with humans, and the target species is anywhere within a class of non-human mammals.

12. 10. The method of claim 1, wherein the one or more comparative data sets comprise from about 10 to about 20 comparative data sets.

13. 1. A system configured for medical image analysis, the system comprising a neural network trained using mammalian transfer learning, the system comprising: (a) receiving one or more comparative datasets, each of the one or more comparative datasets including labeled image data associated with a species; (b) generating a mixed domain dataset based on the one or more comparison datasets; and (c) for each image of the plurality of images of the mixed domain dataset, (i) defining a plurality of chunks within the image, the size of each of the plurality of chunks being selected to obscure a source species of the image; (ii) adding the chunks of the image and any associated labels to a mixed-domain training dataset; (d) using said neural network, (i) training the neural network to discriminate medical properties of case studies from a target species based on the mixed-domain training dataset and the associated labels, wherein the one or more species of the one or more comparison datasets include at least one species other than the target species; (ii) validating the neural network based on a validation data set selected from the mixed-domain data set; (e) generating checkpoints based on the neural network; and 16. A system comprising: a processor configured to:

14. The processor, (a) one or more central processing units in communication with each other directly or via a network; and (b) one or more graphic processing units in communication with each other directly or via a network; The system of claim 13, comprising one or more of:

15. The processor, (a) receiving a case study from a user device, the case study including a set of medical images associated with a patient; (b) defining a second plurality of chunks within the set of medical images, a size of each of the second plurality of chunks being selected to obscure the patient's species; (c) analyzing the second plurality of chunks using the checkpoints to identify the medical characteristic within the set of medical images; (d) providing an indication of said medical characteristic within said set of medical images; and The system of claim 13 , further configured to:

16. When the processor provides the indication of the medical characteristic, (a) displaying at least one medical image of the set of medical images on the user device; (b) displaying a heat map including the indication of the medical characteristic with at least one medical image; The system of claim 15 , further configured to:

17. (a) each of the one or more comparison datasets comprises labeled image data associated with a species and one or more problem classes, the one or more problem classes describing one or more medical abnormalities; (b) the one or more problem classes of the one or more comparison datasets include at least one problem class that does not describe the medical characteristic; The system of claim 13.

18. 14. The system of claim 13, wherein the processor is further configured to test the checkpoint based on a test dataset selected from the mixed domain dataset, the test dataset and the validation dataset each associated with the target species.

19. (a) the one or more comparative datasets comprise a first comparative dataset associated with humans, and the target species is anywhere within a class of mammals other than humans; (b) the one or more comparative datasets comprise from about 10 to about 20 comparative datasets; The system of claim 13.

20. 1. A method for medical image analysis, comprising: (a) providing, by a processor, the case studies to a remote server, (i) the case study is associated with a patient of a target species and comprises a set of medical images associated with the patient; (ii) the remote server is configured to identify medical features of the case studies using a neural network trained by mammalian transfer learning based on a mixed-domain training dataset including labeled image data, the labeled image data comprising: (A) associated with at least one species other than the target species; (B) defining a plurality of chunks, the size of each of the plurality of chunks being selected to obscure a species of source of the labeled image data; Steps and (b) receiving, by the processor, a case study analysis from the remote server, the case study analysis comprising: (i) the set of medical images associated with the patient; and (ii) an analysis dataset describing the medical characteristics exhibited by the set of medical images; and (c) displaying, by said processor and via a display, a visual overlay on at least a portion of said set of medical images, said portion being based on said analysis data set; A method comprising:

Images