Training AI models to ensure balanced performance across subgroups

Abstract

This disclosure describes techniques for facilitating the training of artificial intelligence (AI) models in a manner that ensures balanced model performance across different subgroups. According to one embodiment, a system is provided that includes a memory storing computer-executable components, the processor executing the computer-executable components stored in the memory. The computer-executable components include a training component that trains a machine learning (ML) model on training data using a balanced loss function to perform an inference task, thereby producing a trained version of the ML model. The balanced loss function drives the balanced performance of the ML model across different subgroups represented in the training data, and the trained version provides a defined level of balanced performance across the different subgroups. The balanced loss function is "subgroup-aware" and penalizes variations in model performance across the subgroups during model training and validation.

Description

Technical Field

[0001] This application relates to the development of artificial intelligence (AI) models, and more specifically to training techniques that ensure balanced model performance across different subgroups. Background Technology

[0002] Artificial intelligence (AI) and machine learning (ML) are rapidly evolving technological fields impacting various industries. Advances in machine learning techniques, such as deep neural networks, have recently demonstrated impressive performance in a variety of AI areas, sometimes even surpassing human capabilities. These areas include computer vision, speech, natural language processing (NLP), bioinformatics, drug design, and medical image analysis. As machine learning algorithms continue to evolve to achieve increasingly sophisticated and accurate automated inference capabilities, the goal of ensuring AI's beneficial impact on society has attracted attention in the areas of validation, effectiveness, safety, and control. For example, while AI personal assistants may be annoying, and in some cases laughable, in interpreting NLP errors in everyday commands, the accuracy of AI-based outputs becomes increasingly important when controlling cars, airplanes, automated trading systems, power grids, security systems, and more.

[0003] The performance accuracy of machine learning models / algorithms used in a medical context is critical, especially when the output of machine learning models / algorithms controls the operation of life-sustaining medical devices. Medical software refers to any software or system used in a medical context, such as standalone software for diagnostic or therapeutic purposes, software embedded in medical devices, software that drives medical devices or determines how they are used, software used as accessories to medical devices, software used for the design, production, and testing of medical devices, and software that provides quality control management for medical devices.

[0004] Depending on the type of software associated with it, its intended use, and the level of risk, the regulatory environment for medical software can vary in terms of review requirements and levels. For example, some regulatory agencies require software to be classified as a medical device, often referred to as Software as a Medical Device (SaMD), before it can be authorized for use in clinical practice as a “licensed medical product,” in order to conform to the same regulatory path used for tangible medical devices. Regulatory approval of medical AI models is also conditional on the unbiased performance of their relevant data subgroups. Subgroups are typically constructed using demographic or medical condition metadata. Currently, subgroup analysis is not part of the model training process, and the model’s overall accuracy does not necessarily extend to each of the identified subgroups. Some solutions address this issue by augmenting or increasing sampling to balance the number of training samples from each subgroup. However, these approaches do not guarantee balanced model performance across the domain and may therefore be undesirable for regulatory approval. Summary of the Invention

[0005] The following summary is presented to provide a basic understanding of one or more embodiments of the invention. This summary is not intended to identify key or essential elements, nor is it intended to depict any scope of the different embodiments or any scope of the claims. Its sole purpose is to present concepts in a simplified form as a prelude to the more detailed description that follows. In one or more embodiments, systems, computer-implemented methods, apparatuses, and / or computer program products are described herein that facilitate the development of AI models that perform evenly across different subgroups.

[0006] According to one embodiment, a system is provided including a memory and a processor. The memory stores computer-executable components, and the processor executes the computer-executable components stored in the memory. The computer-executable components include a training component that trains an ML model on training data using a leveling loss function to perform an inference task, thereby producing a trained version of the ML model. The leveling loss function drives the ML model to balance performance across different subgroups represented in the training data, and the trained version provides a defined level of leveling performance across the different subgroups. The different subgroups are labeled in the training data, and the leveling loss function is configured to use the labels to identify the different subgroups. In this respect, the leveling loss function can be considered "subgroup-aware" and penalizes variations in performance across subgroups during model training and validation.

[0007] In some implementations, the elements incorporated in the disclosed system description can be embodied in different forms, such as computer-implemented methods, computer program products, or another form. Attached Figure Description

[0008] Figure 1 A block diagram of an exemplary non-limiting system according to one or more embodiments of the disclosed subject matter is shown, which facilitates the development of AI models that perform evenly across different subgroups.

[0009] Figure 2 The graphs show the performance accuracy of exemplary organ segmentation models trained across different subgroups without a balanced loss function.

[0010] Figure 3 Grouping of training data based on defined subgroups according to one or more embodiments of the disclosed topic is shown.

[0011] Figure 4 An exemplary training process for training an AI model to perform evenly across different subgroups, based on one or more embodiments of the disclosed subject matter, is presented.

[0012] Figure 5A block diagram of an exemplary non-limiting computer implementation of a method according to one or more embodiments of the disclosed subject matter is shown, which facilitates the development of AI models that perform evenly across different subgroups.

[0013] Figure 6 A block diagram of another exemplary non-limiting computer-implemented method according to one or more embodiments of the disclosed subject matter is shown, which facilitates the development of AI models that perform evenly across different subgroups.

[0014] Figure 7 A block diagram is shown illustrating an exemplary, non-limiting operating environment in which one or more embodiments described herein may be facilitated.

[0015] Figure 8 A block diagram is shown of another exemplary non-limiting operating environment in which one or more embodiments described herein may be facilitated. Detailed Implementation

[0016] The following specific embodiments are merely exemplary and are not intended to limit the implementation and / or application or use of the embodiments. Furthermore, they are not intended to be construed as being bound by any express or implied information presented in the foregoing "Background Art" section, "Summary of the Invention" section, or "Detailed Description" section.

[0017] The disclosed subject matter relates to techniques for facilitating the development of AI models that perform evenly across different subgroups. A key challenge proactively addressed by the disclosed techniques is ensuring the development of unbiased AI models that meet regulatory approval requirements. Specifically, bias analysis is targeted at medical AI models by making inferences on demographic and other clinically relevant subgroups, in connection with submissions for regulatory approval (e.g., by the Food and Drug Administration (FDA) and similar agencies). This is typically done after the model has been trained and its biases (if any) have been identified. Defective performance in any identified subgroups leads to further data collection. However, the training and validation processes are unaware of the existence of subgroups, and average accuracy may not extend to all of them.

[0018] To facilitate this purpose, the disclosed subject matter provides AI model training and validation techniques that result in the generation of AI models that perform at a defined subgroup equilibrium (e.g., relative to a defined equilibrium threshold) across the training data. The disclosed techniques take into account the equilibrium performance of the model across different subgroups during a training process using a subgroup-aware loss function (referred to herein as the "equilibrium loss function" or "ELF"). In one or more embodiments, the disclosed techniques group the training data into multiple groups, each group used to identify relevant subgroups. For example, when applied to a clinical AI model suitable for performing clinical inference tasks on medical images or other forms of clinical data, subgroups can be defined based on demographic factors (e.g., patient age, patient sex, image acquisition location, patient body mass index (BMI), etc.), clinical factors (e.g., pathology, comorbidities, etc.), imaging factors (e.g., acquisition protocol, acquisition parameters, etc.), and other factors. The loss across each subgroup can be combined in one of a variety of ways to ensure that the model is trained without bias.

[0019] ELF can be used directly during the model training phase and / or as part of the validation phase to obtain the optimal model with balanced performance across subgroups by penalizing variations in model performance within subgroups and / or weighting the loss inversely to the subgroup sample size. For example, in some implementations, ELF can be configured to compute losses for different subgroups and appropriately weight these losses to drive training optimization during the training phase. Additionally or alternatively, ELF can be configured to compute losses for different subgroups and determine the optimal model based on such losses on the validation dataset. ELF can also be configured to minimize the standard deviation across subgroups. The balance of performance can reflect the same or similar (e.g., relative to a defined degree of similarity) measure of the accuracy and / or specificity of the model performance for each subgroup, measured according to a confidence score, confidence interval, accuracy score, average loss value (e.g., measured using Dice loss, adaptive loss, distance metric, etc.), or another accuracy metric.

[0020] This approach does not require alteration to the training dataset, such as duplication or augmentation, and therefore can be used with any existing training workflow. In this respect, the disclosed technique operates within the available training data and is designed to drive the loss that balances the required model performance levels across defined, relevant subgroups. Therefore, the disclosed technique can ensure compliance with unbiased regulatory requirements for clinical AI models and thus facilitates and simplifies the path to regulatory approval while reducing the amount of time and data required for model training and validation.

[0021] The disclosed techniques can be applied to a variety of fields where AI models are needed or preferred to perform evenly across different subgroups represented in the input data. For example, the disclosed techniques can be applied to the healthcare field, relating to the generation of unbiased clinical inference models considered SaMD and therefore requiring regulatory approval. The disclosed techniques can also be applied to generate unbiased AI / ML models that provide AI informatics at the healthcare administration level to facilitate the planning, regulation, and management of healthcare services. For example, such AI / ML models may include models suitable for processing patient data, relating to optimizing patient scheduling and bed management and improving the clinical and financial outcomes of patient care services. Beyond healthcare, the disclosed techniques can be extended to a variety of other industries or fields. For example, the disclosed techniques can be applied to create unbiased AI / ML models used in the marketing industry to automatically identify trends and deliver more personalized advertising and products to consumers. Other industries that can utilize the disclosed AI / ML model training and development techniques to facilitate the integration of unbiased AI informatics into their systems include, for example, education systems, manufacturing systems, legal systems, personalized assistance systems, government regulatory systems, security systems, machine-to-machine (M2M) communication systems, agricultural systems, and so on. The possibilities are endless.

[0022] Unless the context guarantees a specific distinction between the terms, the terms "algorithm" and "model" are used interchangeably in this document. The terms "AI model" and "ML model" are used interchangeably in this document unless the context guarantees a specific distinction between the terms. The types of AI and ML models or algorithms to which the disclosed unbiased model training and development techniques can be applied can vary. In this regard, specific inference tasks, input data, and model architectures can vary.

[0023] The term "clinical inference model" is used in this document to refer to an AI / ML model configured to perform clinical decision-making / processing on clinical data. Clinical decision-making / processing tasks can vary. For example, clinical decision-making / processing tasks may include classification tasks (e.g., disease classification / diagnosis), disease progression / quantification tasks, organ segmentation tasks, anomaly detection tasks, image reconstruction tasks, etc. Clinical inference models (and other ML models trained according to the disclosed techniques) can employ various types of ML algorithms, including (but not limited to): deep learning models, neural network models, deep neural network models (DNN), convolutional neural network models (CNN), generative adversarial neural network models (GAN), long short-term memory models (LSTM), attention-based models, converters, etc.

[0024] The type of input data processed by the clinical inference model can also vary. For example, in some implementations, the clinical inference model may include a medical image processing model adapted to generate inferences about medical images. As used herein, a “medical imaging inference model” refers to an image inference model customized to perform image processing / analysis tasks on one or more medical images. For example, medical imaging processing / analysis tasks may include (but are not limited to): image reconstruction, image enhancement, scan sequence feature classification, disease / symptom classification, disease segmentation, organ segmentation, disease quantification, disease / symptom staging, risk prediction, temporal analysis, anomaly detection, anatomical feature characterization, medical image reconstruction, etc. Unless the context guarantees a specific distinction between the terms, the terms “medical image inference model,” “medical image processing model,” “medical image analysis model,” etc., are used interchangeably herein.

[0025] The term "image-based inference output" is used herein to refer to a determination or prediction produced by an image processing model configured to perform. For example, image-based inference output may include segmentation masks, reconstructed images, augmented images, adapted images, annotated images, classifications, values, etc. Image-based inference output will vary depending on the type of model and the specific task the model is configured to perform. Image-based inference output may include data objects that can be rendered (e.g., visual data objects), stored, used as input for another processing task, etc. Unless the context guarantees a specific distinction between the terms, the terms "image-based inference output," "inference output," "inference result," "inference," "output," "result," "prediction," etc., are used interchangeably herein. Output may be in various formats, such as: Medicine (DICOM) Structured Report (SR), DICOM secondary capture, DICOM parameter mapping, digital imaging and communication in images, text, and / or JavaScript Object Specification (JSON).

[0026] The types of medical images processed / analyzed by the medical image inference model described herein can include images captured using various types of image capture modalities. For example, medical images can include (but are not limited to): radiotherapy (RT) images, X-ray (XR) images, digital radiography (DX) X-ray images, X-ray angiography (XA) images, panoramic X-ray (PX) images, computed tomography (CT) images, mammography (MG) images (including tomographic synthesis devices), magnetic resonance imaging (MRI) images, ultrasound (US) images, color stream Doppler (CD) images, positron emission tomography (PET) images, single-photon emission computed tomography (SPECT) images, nuclear medicine (NM) images, etc. Medical images can also include synthetic versions of native medical images, such as synthetic X-ray (SXR) images, modified or enhanced versions of native medical images, extended versions of native medical images, and similar versions generated using one or more image processing techniques. The medical imaging processing model disclosed herein can also be configured to process 3D images.

[0027] The disclosed techniques can also be applied to AI / ML models suitable for processing multimodal input data and / or other forms of input data besides images. The term "multimodal data" is used herein to refer to two or more different types of data. The differentiating factors between the two or more different types of data can vary. For example, differentiating factors can refer to the medium of the data (e.g., image data, text data, signal data, etc.), the format of the data, the data capture modality, the source of the data, etc. In a medical / clinical context, multimodal clinical refers to two or more forms of health-related information associated with patient care and / or as part of a clinical trial procedure. Clinical data consists of information ranging from health measurements and assessments of health and health status to documentation of healthcare services. Different types of clinical data are captured for multiple purposes and stored in numerous databases across healthcare systems. Some examples of clinical data that can be processed by clinical inference models may include (but are not limited to): medical images and associated metadata (e.g., acquisition parameters), radiology reports, clinical laboratory data, patient electronic health record (EHR) data, patient physiological data, pharmacy information, pathology reports, admission data, discharge and transfer data, discharge summaries, progress notes, demographic data, and other non-imaging data.

[0028] One or more embodiments will now be described with reference to the accompanying drawings, wherein the same reference numerals are always used to denote the same elements. In the following description, numerous specific details are set forth for purposes of explanation in order to provide a more thorough understanding of one or more embodiments. However, it will be apparent, in various cases, that one or more embodiments may be practiced without these specific details.

[0029] Figure 1 A block diagram of an exemplary model development system 100 according to one or more embodiments of the disclosed subject matter is shown, which facilitates the development of AI models that perform in a balanced manner across different subgroups. Embodiments of the systems described herein may include one or more machine-executable components embodied within one or more machines (e.g., embodied in one or more computer-readable storage media associated with one or more machines). When executed by one or more machines (e.g., processors, computers, computing devices, virtual machines, etc.), such components can enable the one or more machines to perform the operations described herein.

[0030] In this regard, the model development system 100 includes a training component 102, a machine learning (ML) model 104, a training data management component 106, an equalization loss function (ELF) 108, and a reporting component 110, all of which may be or are included in machine-executable components embodied within one or more machines (e.g., embodied in one or more computer-readable storage media associated with one or more machines), which, when executed by one or more machines (e.g., processors, computers, computing devices, virtual machines, etc.), enable one or more machines to perform the described operations. For example, the model development system 100 may be any suitable machine capable of performing one or more of the operations described by reference to the training component 102, the ML model 104, the training data management component 106, the ELF 108, and the reporting component 110.

[0031] As used herein, a machine can be and / or may include one or more of the following: a computing device, a general-purpose computer, a special-purpose computer, a quantum computing device (e.g., a quantum computer), a tablet computing device, a handheld device, a server-type computer and / or a database, a laptop computer, a notebook computer, a desktop computer, a mobile phone, a smartphone, a consumer device and / or instrumentation, an industrial and / or commercial device, a digital assistant, a multimedia internet-enabled phone, and / or another type of device. The model development system 100 may also be or correspond to one or more physical or virtual (e.g., cloud-based) computing devices. The model development system 100 may also include or be operatively coupled to at least one memory 114 that stores computer-executable components (e.g., training component 102, ML model 104, training data management component 106, ELF 108, and reporting component 110). Memory 114 may also store training data 116 and / or training data 116 may be stored at another suitable data storage location accessible to the model development system 100 (e.g., via one or more networks). The model development system 100 may also include, or be operatively coupled to, at least one processing unit 112 (or processor) that executes computer-executable components stored in memory 114. Examples of the memory 114 and processing unit 112, as well as other suitable computer or computing-based elements, may be found in [reference needed]. Figure 7 (For example, refer to processing unit 704 and system memory 706) to find, and can be combined with implementation Figure 1 It may be used in accordance with one or more of the systems or components shown and described in the other accompanying drawings disclosed herein.

[0032] The deployment architecture of the model development system 100 is variable. In some implementations, the model development system 100 can be deployed as a local computing device. In various implementations, one or more of the components of system 100 can be deployed in a cloud architecture, a virtualized enterprise architecture, or an enterprise architecture, wherein one of the front-end components and the back-end components is distributed in a client / server relationship. Using these implementations, the features and functions of one or more of the training component 102, ML model 104, training data management component 106, ELF 108, reporting component 110, processing unit 112, and / or memory can be deployed as a web application, cloud application, thin client application, thick client application, local client application, hybrid client application, etc. See below for reference. Figures 7 to 8 Various exemplary deployment architectures for system 100 are described.

[0033] Training component 102 provides training for ML model 104 to perform defined inference tasks on training data 116 using ELF 108, wherein ELF 108 drives balanced performance of ML model 104 across different subgroups represented in training data 118. The specific type of ML model 104, the inference task performed by ML model 104, the input data processed by ML model 104 (i.e., training data 116), and the factors defining the different subgroups can vary. For example, in some embodiments, ML model 104 may include or correspond to a clinical inference model suitable for performing clinical inference tasks on clinical data. Using these embodiments, subgroups can be defined based on clinical and / or non-clinical factors associated with the clinical data. For example, assuming training data 116 includes clinical data associated with different patients, such as medical images acquired for different patients and / or other forms of clinical and non-clinical data associated with different patients (e.g., laboratory data, pathology data, demographic data, acquisition protocol data, etc.), subgroups can be defined based on different types of patients and / or patient groups by one or more factors. For example, different types of patients or patient groups can be defined based on demographic attributes (e.g., age, sex, body mass index (BMI), height / weight, location), clinical factors (e.g., comorbidities, risk level, pathology, etc.), medical history factors, psychosocial risk factors, etc.

[0034] The training process performed by training component 102 may involve the conventional ML training and validation and testing phases associated with supervised, semi-supervised, and / or unsupervised ML training and development processes, where ELF 108 is used instead of the conventional loss function. A detailed description of the training process and ELF 108 is described below. This training process results in the generation of trained versions of the ML model 104 that provide a defined, balanced performance level across different subgroups. This performance balance may reflect identical or similar (e.g., relative to a defined degree of similarity) measures of accuracy and / or specificity of the model performance for each subgroup, measured according to a confidence score, confidence interval, accuracy score, average loss value (e.g., measured using Dice loss, adaptive loss, distance metric, etc.), or another accuracy metric.

[0035] Figure 2Plot 200 is presented, illustrating the performance accuracy of an exemplary organ segmentation model trained across different subgroups without an equalization loss function (such as ELF 108). The organ segmentation model corresponds to a liver segmentation model trained on ultrasound medical images captured in the abdomen to segment the liver (e.g., via predictions of corresponding segmentation masks on the liver). The liver segmentation model employs the UNet architecture and was developed using a standard supervised ML training and validation / testing process. The training data comprised medical images captured for approximately 400 subjects (or patients) divided into training, validation, and testing subsets. Subjects were distinguished by several clinically relevant factors, including those that constitute the subgroups identified in Plot 200. However, these subgroups were not labeled or evaluated during the model training and validation / testing phases. Instead, after model training, images in the validation set were grouped into distinct subgroups based on the factors identified in Plot 200. Specifically, images were grouped by country of origin (e.g., including China, the United States, Poland, and the Czech Republic), male and female, age less than 60, age greater than 60, BMI less than 25, and BMI greater than 25. These different subgroups were clinically determined to be relevant by experts. The total number of images included in the validation set was 133, and the corresponding number of images included within each subgroup was indicated by the subgroup title immediately following it.

[0036] The model's average accuracy relative to all data samples (i.e., images) included in the validation set is further determined based on the confidence interval, which represents the confidence level in the model's ability to generate accurate liver segmentation mask predictions. In this example, the 90% confidence interval marks the minimum acceptable confidence level required for all subgroups. The vertical lines at the top of each column indicate the upper and lower CI boundaries determined for each group.

[0037] As shown in Figure 200, although the mean CI for each subgroup is above the 90% minimum, the mean CI varies across subgroups, thus indicating that model performance accuracy varies across subgroups. For example, the mean model CI for images acquired in China is significantly higher than that for images acquired in Poland. Similarly, the mean CI for images captured from male subjects is higher than that for female subjects. Figure 200 also demonstrates that the lower CI limit for all subjects acquired jointly is at or near the 95% mark; however, the lower CI limit for several subgroups drops below the 90% mark. Therefore, as demonstrated by Figure 200, conventional model development techniques cannot guarantee balanced model performance across different clinically relevant subgroups. Typically, to address this issue, additional data samples are acquired for the poorly performing subgroups, and the model is retrained and re-evaluated using these additional data samples until model accuracy becomes homogeneous across all subgroups (e.g., reaching a defined maximum level of bias). However, such techniques are limited by the available training data, require additional training time, and cannot guarantee model accuracy across different subgroups after field deployment.

[0038] Refer again Figure 1 In light of these shortcomings, the disclosed ML model training technique explicitly requires that the ML model 104 perform uniformly across all qualified subgroups during training using ELF 108. In this regard, ELF 108 corresponds to a loss function specifically configured to know the existence of qualified subgroups of the data samples represented in training data 116 and drive the ML model 104 to achieve balanced performance across different subgroups during the model training and / or validation phases. To facilitate this, the corresponding samples in training data 116 can be associated with information identifying the qualified subgroups (or subgroups) to which they belong. Such information can be associated with the data samples as annotation tags, metadata (e.g., metadata tags), or in a separate indexed data structure. Specific differentiation factors that qualify the relevant subgroups represented in training data 116 can be predefined. In some implementations, ELF 108 can be configured to use this information directly during the training phase to identify the specific subgroup to which a data sample belongs and accordingly weigh the loss, as described in more detail below.

[0039] Additional or alternative land, such as Figure 3As shown, the training data management component 106 can divide the training data 116 into a training set 301, a validation set 302, and a test set 303. In some implementations, the validation set 302 and the test set 303 can be combined. In this regard, the training set 301 is used during the model training phase to fit the ML model 104, according to conventional ML training techniques. The validation set 302 is used to provide an unbiased evaluation of the model fitted on the training set 301, while tuning the model hyperparameters. As the techniques on the validation set 302 are incorporated into the model configuration, the evaluation becomes more biased. In this regard, the validation loss is used to select the best version of the ML model 104 generated during the training phase and to avoid overfitting. The test set 303 is used to measure the performance of the version of the ML model 104 selected during the validation phase.

[0040] The training data management component 106 can further group data samples contained in the training set 301, validation set 302, and test set 303 into different subgroups based on predefined grouping factors (e.g., demographic factors, pathological factors, image acquisition factors, etc.). Information defining the predefined grouping factors can be stored in memory 114, and the training data management component 106 can perform grouping using metadata tags associated with the training data samples contained in the training data 116. In other embodiments, the training data management component 106 can employ one or more machine learning techniques to automatically cluster the training data 116 into two or more distinct subgroups, such as k-means clustering, hierarchical clustering, etc. Using these embodiments, the training data management component 106 can infer grouping factors. Figure 3 In the example shown, the subgroups include five distinct subgroups (identified as group 1, group 2, group 3, group 4, and group 5, respectively). The factors used to define the five subgroups are arbitrary for the purposes of this example, as they will depend on the type of ML model 104, the task of ML model 104, the type of training data, etc. Similarly, it should be understood that the number of different subgroups can vary. The number of data samples included in the training set 301, validation set 302, and test set 303, respectively, can also vary. Typically, the training set 301 may contain approximately 60% of the samples included in the training data 116, and the validation set 302 and test set 303 may each contain approximately 20% of the data samples; however, these distributions can vary. In most implementations, the training set 301 includes at least 50% of the data samples included in the training data.

[0041] The distribution of data samples included in the corresponding subgroups for each of the training set 301, validation set 302, and test set 303 can also vary and will depend on the distribution of the corresponding subgroups included in the training data 116. In some embodiments, the training data management component 106 may randomly divide the training data 116 into training set 301, validation set 302, and test set 303, and then cluster the corresponding sets into different subgroups. Additionally or alternatively, the training data management component 106 may perform class balancing to balance the number of data samples contained in each group within each of the training set 301, validation set 302, and test set 303 according to the defined data partitioning for the training set 301, validation set 302, and test set. For example, assuming the corresponding sets are partitioned according to a 60:20:20 ratio, the distribution of data samples included in group 1 can also be partitioned among the training set 301, validation set 302, and test set 303 according to a 60:20:20 ratio. For example, in some implementations, the training data management component 106 can initially divide the data samples contained in the training data 116 into defined subgroups (e.g., Figure 3 (As shown in Group 1, Group 2, Group 3, Group 4, and Group 5). Subsequently, the training data management component 106 can distribute each subgroup of data samples between the training set 301, validation set 302, and test set 303 according to a defined ratio (e.g., 60:20:20, 60:25:15, 50:25:25, 70:15:15, or another ratio).

[0042] It should be understood that in various implementations, subgroups are not mutually exclusive. For example, according to Figure 1 The liver segmentation model subgroups shown are divided by geographic region, sex, age, and BMI. Naturally, those data samples (e.g., abdominal ultrasound images) included in any of the geographic subgroups can include data samples belonging to male and female subjects, subjects aged 60 years and older, and subjects with a BMI of 25 or higher. Therefore, in some embodiments, within the training set 301, validation set 302, and / or test set 303, the training data management component 106 can generate multiple instances of data samples belonging to two or more groups, and generate an associated instance within each group to which it belongs.

[0043] In various implementations, besides the use of ELF 108, ML model 104 includes a supervised or semi-supervised ML model, and training component 102 can train ML model 104 using a conventional supervised or semi-supervised machine learning process. In this regard, within the supervised learning framework, ML model 104 is first given examples of data from training data 116 where human experts or annotators apply ground-true classification labels or other forms of ground-true annotations (e.g., segmentation masks, bounding boxes, etc.) to some or all samples included. ML then uses the ground-true annotations to adjust and modify the internal mathematical representation of the data (such as the behavior of an artificial neural network) and map it to certain predictions for classification, etc. Training consists of an iterative method using numerical optimization techniques that reduce the error or loss between the desired class label and the ML model's predictions. New data is then fed into the newly trained model as input, and if the training is good, the new data can be classified or otherwise evaluated.

[0044] According to the disclosed technology, training component 102 can use ELF 108 to determine the loss between corresponding training data samples and their ground truth examples, which ELF 108 computes the training and / or validation loss for each of the defined subgroups. Training component 102 can further combine the losses across each subgroup in one of a variety of ways to ensure that the ML model 104 is trained without bias. For example, in some embodiments, training component 102 can use ELF 108 to compute the losses for different subgroups during the model training phase and appropriately weigh these losses to drive training optimization during the training phase. Additionally or alternatively, ELF can be configured to compute the losses for different subgroups and determine the optimal model based on such losses on the validation dataset.

[0045] Specifically, ELF 108 can be configured to identify the specific subgroup to which each data sample belongs as information previously associated with the corresponding data samples that identified it to belong to one or more subgroups. Such information can be associated with the corresponding data sample as annotation tags, as metadata tags in metadata, or in another form. ELF 108 can further penalize performance variations across subgroups. In this regard, the difference between training loss and / or validation loss across subgroups can be considered in ELF 108 to achieve balanced performance of the ML model 104 across subgroups. In some implementations, if certain subgroups have lower representations in the population, ELF 108 can be configured to be more stringent towards these groups to achieve tighter confidence intervals. In this regard, the way ELF 108 is configured to penalize performance variations across subgroups based on loss can vary.

[0046] In one exemplary embodiment, ELF 108 may include a loss function configured to weigh the loss for different subgroups (e.g., measured according to Dice loss, overlap loss, or another loss metric) relative to the number of samples in each subgroup contained within the entire training data 116, such that more weight is given to subgroups that are not adequately represented. In this respect, the loss weights defined for the respective subgroups may be proportional to the amount of representation in the training data 116, where the smaller the number of data samples included in the subgroup, the higher the loss weight. In this way, the loss associated with subgroups that are not adequately represented is amplified to balance the loss associated with subgroups with higher representation. According to this embodiment, ELF 108 may be defined according to Equation 1, where is the weight associated with the i-th subgroup, and is the loss associated with it, where the weights may be inversely proportional to the sample size of the subgroup. Using this embodiment, the loss is biased to provide higher accuracy for subgroups with smaller sample sizes, thereby tightening the CI bound of the model.

[0047]

[0048] Equation 1

[0049] In another exemplary embodiment, ELF 108 may include a loss function configured to explicitly reduce the variance of the loss across subgroups to ensure that the loss (e.g., measured according to Dice loss, overlap loss, or another loss metric) is the same or substantially the same across different subgroups (e.g., relative to a defined maximum deviation threshold). In other words, ELF 108 may be configured to minimize the standard deviation of the loss across subgroups. Using these embodiments, ELF 108 can be defined according to Equation 2 below, where is the loss associated with the subgroup and is the average loss across all subgroups. According to Equation 2, both the total loss and variance are minimized and biased to have the same accuracy across all subgroups (relative to a defined degree of similarity for the accuracy measure).

[0050]

[0051] Equation 2.

[0052] It should be understood that the exemplary equalization loss function defined by Equations 1 and 2 is merely exemplary, and other mechanisms are envisioned for configuring ELF 108 to penalize variations in model performance across subgroups.

[0053] As described above, training component 102 may employ ELF 108 corresponding to Equations 1, 2, or similar equations during the training and / or validation phases of the ML model 104 training and development process. During these respective phases, ELF 108 calculates the loss for the respective subgroup. Training component 102 may further employ the loss calculated for the respective subgroup to generate an accuracy measurement in the model for the respective subgroup based on the corresponding loss. In some embodiments, upon completion of training (e.g., the training and validation phases), training component 102 may use a test set that has been divided into different subgroups to determine a final measurement of the accuracy of the ML model's performance for each subgroup within the subgroups. For example, training component 102 may determine confidence scores and / or confidence intervals for the overall model and for each subgroup within the respective subgroups based on the average, upper, and lower loss values ​​associated with each subgroup within the subgroups determined during the training, validation, and / or testing phases, wherein the confidence scores and / or confidence intervals represent a confidence level of the model's accuracy on each subgroup within the subgroups. Reporting component 110 can further generate a report identifying subgroup-specific performance accuracy information. This may include subgroup-specific model performance accuracy information associated with the submission of ML model 104 for regulatory approval, used to measure changes in model performance within the field and / or for downstream processing tasks. Since ELF 108 is configured to drive models with identical performance (or substantially identical relative to a defined degree of deviation) for each subgroup, accuracy measurements in the models for the corresponding subgroups will be identical (or substantially identical). For example, refer again... Figure 2 The training component 102 can determine for each subgroup in the subgroup. Figure 2 The confidence interval information shown (e.g., including average CI and upper and lower limits) is provided, and the reporting component 110 can generate a report that includes the confidence interval information shown in graph 200. However, minimizing the differences between subgroups is also possible. Figure 2 The variance is shown in the figure. In this respect, the confidence interval information report generated for the ML model trained using the disclosed technique can be used to confirm the model's balanced performance across subgroups.

[0054] Figure 4 An exemplary training process 400 for training an ML model 104 to perform evenly across different subgroups, according to one or more embodiments of the disclosed subject matter, is presented. In various embodiments, training component 102 may apply process 400 to train the ML model 104 to perform evenly across the defined subgroups represented in training data 116, thereby producing a trained version of the ML model 104 that is unbiased towards the different subgroups upon completion of the training and validation phases. For brevity, repeated descriptions of similar elements employed in the respective embodiments are omitted.

[0055] Training process 400 may initially involve preprocessing training data 116 at 402 to partition training data 116 into training set 404 and validation set 406 (e.g., via training data management component 106). Preprocessing at 402 may also involve grouping the corresponding data samples contained in training set 404 and / or validation set 406 into defined subgroups using information associated with corresponding samples identifying to which they belong (e.g., metadata tags, etc.) (e.g., via management component 106). In other embodiments, ELF 108 may implicitly perform this grouping, associated with calculating the loss using subgroup tags / metadata tags associated with the corresponding samples. In this example, training data 116 is not partitioned into a third test set; however, it should be understood that in other embodiments, training data 116 may also be partitioned into a test set. Nevertheless, training process 400 illustrates the training and validation phases of the model training and development process. It should be understood that the final model performance after process 400 is completed can be evaluated at the subgroup level using test sets that have also been subdivided into defined relevant subgroups. During the testing phase, the mode configuration and parameters are not adjusted. Preprocessing at 402 may also involve data cleaning / engineering to fill out outliers and / or missing values.

[0056] According to procedure 408, training set 404 is used to train and build ML model 104. Procedure 408 involves model building at 410, which involves applying input data samples contained in training set 404 to ML model 104 to generate training result 412. For example, assuming ML model 104 includes a liver segmentation model and the input data samples include medical images of the abdomen, training result 412 will include the corresponding liver segmentation mask. At 414, procedure 408 involves determining and evaluating the loss using ELF 108 based on the ground truth (GT) annotations applied to the corresponding input data samples. As mentioned above, the loss function evaluation at 414 typically involves using ELF 108 to determine a measure of the accuracy difference between training result 412 and the actual GT. The measure of the accuracy difference can be based on distance metrics such as Hamming distance, Jaccard distance / exponent, Dice score / coefficient, adaptive loss, etc.

[0057] Based on the subject training phase used by ELF 108 to determine loss, ELF 108 can be configured to evaluate loss based on subgroup classification for each data sample. For example, in some implementations, ELF 108 can be configured according to Equation 1 and use defined weights determined for each subgroup in the different subgroups to weigh the loss, wherein the weights are inversely proportional to the number of data samples contained in each subgroup (e.g., indicating that the smaller the number, the higher the weight). In some implementations of these embodiments, the preprocessing at 402 may also involve determining the relative distribution of the data samples included in each subgroup (e.g., relative to the training set 404 and / or the training data 116 as a whole) and defining the weights to be used for ELF 108 for each subgroup based on the relative distribution.

[0058] At 416, process 408 further includes adjusting model weights and / or parameters based on the loss to reduce the amount of loss on the next evaluation of the next training data sample. This process 408 is performed iteratively until the model loss stabilizes to a defined extent and / or otherwise converges.

[0059] Once the ML model has progressed to a point of sufficiency on training set 404 (e.g., until the model loss has stabilized to a defined extent and / or otherwise converges), training component 102 can perform a validation phase of the model training process using validation set 406. In this regard, training component 102 can apply validation set 406 to ML model 104 to generate validation results 418. Training component 102 can further evaluate validation results 418 at 420 using ELF 108 to assess the performance accuracy and specificity of ML model 104. For example, training component 102 can use ELF 108 to determine the loss across validation results 418 and the corresponding ground truth paradigms for the data samples included in validation set 406. Related to determining the loss, training component 102 can use ELF 108 to determine the loss for different subgroups defined in the validation set. ELF 108 can further penalize variations in performance across subgroups to drive balanced performance of ML model 104 across subgroups. For example, the loss can be interpreted to minimize the standard deviation across subgroups. The version of ELF 108 used at 420 may correspond to the same or a different version of ELF 108 used during the training phase and process 408. For example, in some implementations, the version of ELF 108 used at 420 may correspond to Equation 2. Using these implementations, training component 102 may compute the loss for each of the defined subgroups represented in validation set 406 and determine the optimal version of ML model 104 tuned during training phase 408 that minimizes the total loss and variance across subgroups (e.g., minimizes the standard deviation across subgroups to a defined degree of bias). Once the model training and validation phases have been completed, ML model 104 may optionally be tested using a new (unseen) test dataset that includes data samples from different subgroups.

[0060] Figure 5 A block diagram of an exemplary non-limiting computer implementation of method 500 according to one or more embodiments of the disclosed subject matter is shown, which facilitates the development of AI models that perform evenly across different subgroups. For brevity, repeated descriptions of similar elements employed in the respective embodiments are omitted.

[0061] According to method 500, at 502, a system operatively coupled to a processor (e.g., system 100, etc.) trains (e.g., using training component 102) a machine learning model (e.g., ML model 104) on training data (e.g., training data 116) using an equalization loss function (e.g., ELF 108) to perform an inference task. This equalization loss function drives the equalization performance of the machine learning model across different subgroups represented in the training data. At 504, the system generates a trained version of the machine learning model based on the training, which provides a defined level of equalization performance across the different subgroups (e.g., via training component 102).

[0062] Figure 6 A block diagram of another exemplary non-limiting computer implementation of a method 600 according to one or more embodiments of the disclosed subject matter is shown, which facilitates the development of AI models that perform evenly across different subgroups. For brevity, repeated descriptions of similar elements employed in the respective embodiments are omitted.

[0063] According to method 600, at 602, a system operatively coupled to the processor (e.g., system 100, etc.) groups the training dataset and validation dataset into subgroups comprising different subgroups of data samples (e.g., using training data management unit 106). At 604, the system uses the training dataset during the training phase to train (e.g., via training unit 102) a machine learning model to perform tasks using a first equilibrium loss function (e.g., ELF 108), which drives the machine learning model to balance performance across the different subgroups. At 606, the system may further use the validation dataset during the validation phase to determine (e.g., via training unit 102) the optimal configuration of the machine learning model generated during the training phase using a second equilibrium loss function (e.g., ELF 108) that minimizes the loss between subgroups. In some embodiments, the first equilibrium loss function and the second equilibrium loss function are the same. In other embodiments, the first equilibrium loss function and the second equilibrium loss function may be different (e.g., Equations 1 and 2). In other embodiments, the equilibrium loss function may be used during either the training phase or the validation phase, but not both.

[0064] One or more embodiments may be systems, methods, and / or computer program products at any possible level of technical detail in integration. A computer program product may include a computer-readable storage medium (or media) having computer-readable program instructions thereon for causing a processor to perform one or more aspects of this embodiment.

[0065] Computer-readable storage media can be tangible devices that hold and store instructions for use by instruction execution devices. Computer-readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of computer-readable storage media includes: portable computer floppy disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable optical disc read-only memory (CD-ROM), digital versatile optical disc (DVD), memory sticks, floppy disks, mechanical encoding devices (such as punch cards or raised structures in recesses on which instructions are recorded), and any suitable combination of the foregoing. As used herein, computer-readable storage media should not be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses passing through fiber optic cables), or electrical signals transmitted through wires.

[0066] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to a suitable computing / processing device, or downloaded via a network (e.g., the Internet, a local area network, a wide area network, and / or a wireless network) to an external computer or external storage device. The network may include copper cables, optical fibers, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them for use in a computer-readable storage medium stored within the suitable computing / processing device.

[0067] Computer-readable program instructions used to perform the operations of this invention may be assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​(including object-oriented programming languages ​​such as Smalltalk, C++, etc.) and procedural programming languages ​​(such as the "C" programming language or similar programming languages). The computer-readable program instructions may be executed entirely on a physical computer, partially on a physical computer, as a standalone software package, partially on a physical computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the physical computer via any type of network (including local area network (LAN) or wide area network (WAN)) or may be connected to an external computer (e.g., via the Internet through an Internet service provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) may execute the computer-readable program instructions to personalize the electronic circuitry in order to perform aspects of this invention by utilizing the state information of the computer-readable program instructions.

[0068] This document describes aspects of the invention with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0069] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer-readable storage medium having the instructions stored therein includes an article of writing comprising instructions that implement aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram.

[0070] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer-implemented process, such that the instructions executed on the computer, other programmable apparatus or other device implement the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0071] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible specific implementations of systems, methods, and computer program products according to various embodiments of the invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of instructions comprising one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions indicated in the blocks may not occur in the order shown in the figures. For example, depending on the function involved, two blocks shown consecutively may actually be executed substantially simultaneously, or sometimes they may be executed in reverse order. It will also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified function or action or implements a combination of dedicated hardware and computer instructions.

[0072] Combination Figure 7 The systems and processes described below may be embodied in hardware, such as a single integrated circuit (IC) chip, multiple ICs, application-specific integrated circuits (ASICs), etc. Furthermore, the order in which some or all program blocks appear in each program should not be considered restrictive. Rather, it should be understood that some program blocks can be executed in various orders, and not all of the orders described can be explicitly shown herein.

[0073] refer to Figure 7 An example environment 700 for implementing various aspects of the claimed subject matter includes a computer 702. The computer 702 includes a processing unit 704, system memory 706, a codec 735, and a system bus 708. The system bus 708 couples system components, including but not limited to system memory 706, to the processing unit 704. The processing unit 704 can be any of a variety of available processors. Dual microprocessors and other multiprocessor architectures may also be used as the processing unit 704.

[0074] The system bus 708 can be any of a variety of bus architectures, including memory bus or memory controller, peripheral bus or external bus, or local bus using any of the available bus architectures, including but not limited to Industry Standard Architecture (ISA), Micro Channel Architecture (MSA), Extended ISA (EISA), Intelligent Drive Electronic Devices (IDE), VESA Local Bus (VLB), Peripheral Component Interconnect (PCI), Card Bus, Universal Serial Bus (USB), Advanced Graphics Port (AGP), PCMCIA Bus, FireWire (IEEE 1374), and Small Computer System Interface (SCSI).

[0075] In various embodiments, system memory 706 includes volatile memory 710 and non-volatile memory 712, which may employ one or more memory architectures among the disclosed memory architectures. A basic input / output system (BIOS) containing basic routines such as transferring information between elements within computer 702 during startup is stored in non-volatile memory 712. Additionally, according to the invention, codec 735 may include at least one of an encoder or a decoder, wherein the encoder or decoder may be composed of hardware, software, or a combination of hardware and software. Although codec 735 is depicted as a separate component, codec 735 may be contained within non-volatile memory 712. By way of illustration and not limitation, non-volatile memory 712 may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, 3D flash memory, or resistive memory, such as resistive random access memory (RRAM). In at least some embodiments, non-volatile memory 712 may employ one or more memory devices among the disclosed memory devices. Furthermore, the non-volatile memory 712 may be computer memory (e.g., physically integrated with computer 702 or its motherboard) or removable memory. Examples of suitable removable memory that can be used to implement the disclosed embodiments may include a Secure Digital (SD) card, a Compact Flash (CF) card, a Universal Serial Bus (USB) Memory Stick, etc. Volatile memory 710 includes random access memory (RAM) that acts as external cache memory, and one or more of the disclosed memory devices may also be employed in various embodiments. By way of illustration and not limitation, RAM can be provided in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and enhanced SDRAM (ESDRAM), etc.

[0076] Computer 702 may also include removable / non-removable, volatile / non-volatile computer storage media. Figure 7Disk storage device 714 is illustrated, for example. Disk storage device 714 includes, but is not limited to, devices such as disk drives, solid-state drives (SSDs), flash memory cards, or memory sticks. Additionally, disk storage device 714 may include storage media alone or in combination with other storage media, including but not limited to optical disc drives, such as optical disc ROM devices (CD-ROM), CD recordable drives (CD-R drives), CD rewritable drives (CD-RW drives), or digital versatile optical disc ROM drives (DVD-ROM). To facilitate connection of disk storage device 714 to system bus 708, removable or non-removable interfaces, such as interface 716, are typically used. It should be understood that disk storage device 714 may store information relating to an entity. Such information may be stored at a server or provided to applications running on the server or entity device. In one embodiment, the type of information stored on disk storage device 714 or transmitted to the server or application may be notified to the entity (e.g., via one or more output devices 736). Entities may be provided with the opportunity to opt in or out of using the server or application to collect or share such information (e.g., via input from input device 728).

[0077] It should be understood that Figure 7 The description refers to software that acts as an intermediary between an entity and the basic computer resources described in the suitable operating environment 700. This software includes an operating system 718. The operating system 718, which may be stored on a disk storage device 714, is used to control and allocate the resources of the computer system 702. Application programs 720 utilize the operating system 718 for resource management through program modules 724, and program data 726 stored in system memory 706 or on disk storage device 714, such as power-on / power-off transaction tables. It should be understood that the claimed subject matter can be implemented using various operating systems or combinations of operating systems.

[0078] An entity inputs commands or information into computer 702 via input device 728. Input device 728 includes, but is not limited to, pointing devices such as mouse, trackball, stylus, touchpad, keyboard, microphone, joystick, gamepad, satellite dish, scanner, TV tuner card, digital camera, digital camcorder, webcam, etc. These and other input devices are connected to processing unit 704 via interface port 730 through system bus 708. Interface port 730 includes, for example, serial port, parallel port, game port, and Universal Serial Bus (USB). Output device 736 uses some of the same type of ports as input device 728. Thus, for example, a USB port can be used to provide input to computer 702 and output information from computer 702 to output device 736. Output adapter 734 is provided to illustrate some output devices 736 such as monitors, speakers, and printers, as well as other output devices 736 that require special adapters. By way of illustration and not limitation, output adapter 734 includes video cards and sound cards that provide a means of connection between output device 736 and system bus 708. It should be noted that other devices or systems of devices provide both input and output capabilities, such as the remote computer 738.

[0079] Computer 702 can operate in a networked environment using a logical connection to one or more remote computers, such as remote computer 738. Remote computer 738 can be a personal computer, server, router, network PC, workstation, microprocessor-based device, peer-to-peer device, smartphone, tablet, or other network node, and typically includes many of the elements described with respect to computer 702. For simplicity, only memory storage device 740 is shown for remote computer 738. Remote computer 738 is logically connected to computer 702 via network interface 742 and subsequently connected via communication connection 744. Network interface 742 covers wired or wireless communication networks, such as local area networks (LANs) and wide area networks (WANs), as well as cellular networks. LAN technologies include Fiber Distributed Data Interface (FDDI), Copper Distributed Data Interface (CDDI), Ethernet, Token Ring, etc. WAN technologies include, but are not limited to, point-to-point links, such as Integrated Services Digital Network (ISDN) and its variants, circuit-switched networks, packet-switched networks, and Digital Subscriber Line (DSL).

[0080] Communication connection 744 refers to the hardware / software used to connect network interface 742 to bus 708. Although communication connection 744 is shown inside computer 702 for clarity, the communication connection may also be outside computer 702. For illustrative purposes only, the hardware / software required to connect to network interface 742 includes internal and external technologies such as modems, including conventional telephone-grade modems, cable modems and DSL modems, ISDN adapters, and wired and wireless Ethernet cards, hubs and routers.

[0081] The aspects illustrated in this disclosure can also be practiced in a distributed computing environment, where certain tasks are performed by remote processing devices linked via a communication network. In a distributed computing environment, program modules can reside in both local and remote memory storage devices.

[0082] refer to Figure 8 This diagram illustrates a schematic block diagram of a computing environment 800 according to the present disclosure, in which subject systems (e.g., system 100, etc.), methods, and computer-readable media may be deployed. The computing environment 800 includes one or more clients 802 (e.g., laptop computers, smartphones, PDAs, media players, computers, portable electronic devices, tablets, etc.). Clients 802 may be hardware and / or software (e.g., threads, processes, computing devices). The computing environment 800 also includes one or more servers 804. Servers 804 may also be hardware or hardware combined with software (e.g., threads, processes, computing devices). For example, server 804 may accommodate threads to perform transformations by employing aspects of the present disclosure. In various embodiments, one or more components, devices, systems, or subsystems of system 100 may be deployed as hardware and / or software at client 802 and / or deployed as hardware and / or software at server 804. One possible form of communication between client 802 and server 804 may be in the form of data packets transmitted between two or more computer processes. These data packets may include healthcare-related data, training data, AI models, input data to the AI ​​models, encrypted output data generated by the AI ​​models, etc. For example, the data packets may include metadata, such as associated contextual information. Computing environment 800 includes a communication framework 806 (e.g., a global communication network such as the Internet, or a mobile network) that can be used to facilitate communication between client 802 and server 804.

[0083] Communication can be facilitated via wired (including fiber optic) and / or wireless technologies. Client 802 includes or is operatively connected to one or more client data storage devices 808 that can be used to store information local to client 802 (e.g., associated context information). Similarly, server 804 operatively includes or is operatively connected to one or more server data storage devices 810 that can be used to store information local to server 804 (e.g., ML model 104 and its trained unbiased version, ELF 108, training data 116, etc.).

[0084] In one implementation, client 802 may transmit encoded files to server 804 according to the disclosed subject matter. Server 804 may store files, decode files, or transmit files to another client 802. It should be understood that client 802 may also transfer uncompressed files to server 804 and may compress files according to the disclosed subject matter. Similarly, server 804 may encode video information and transmit the information to one or more clients 802 via communication framework 806.

[0085] Although the subject matter has been described above in the general context of computer executable instructions for a computer program product running on one or more computers, those skilled in the art will recognize that the present disclosure may also be implemented in combination with other program modules. Generally, program modules include routines, programs, components, data structures, etc., that perform specific tasks and / or implement specific abstract data types. Furthermore, those skilled in the art will understand that the computer implementation of the methods of the present invention can be practiced using other computer system configurations, including single-processor or multi-processor computer systems, small computing devices, mainframe computers, and computers, handheld computing devices (e.g., PDAs, telephones), microprocessor-based or programmable consumer or industrial electronic devices, etc. The illustrated aspects can also be practiced in a distributed computing environment where tasks are performed by remote processing devices linked via a communication network. However, some (if not all) aspects of the present disclosure can be practiced on a standalone computer. In a distributed computing environment, program modules can reside in both local and remote memory storage devices.

[0086] As used herein, the terms “component,” “system,” “subsystem,” “platform,” “layer,” “gateway,” “interface,” “service,” “application,” “device,” etc., may refer to and / or include one or more computer-related entities or entities associated with an operating machine having one or more specific functions. Entities disclosed herein may be hardware, a combination of hardware and software, software, or software in execution. For example, a component may be, but is not limited to, a program running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. By way of illustration, an application running on a server and a server may both be components. One or more components may reside within a program and / or an execution thread, and components may be located on a single computer and / or distributed across two or more computers. In another example, a corresponding component may be executable on various computer-readable media on which various data structures are stored. Components may communicate via local and / or remote processes, such as based on signals having one or more data packets (e.g., data from one component that interacts with another component in a local system, a distributed system, and / or a network (such as the Internet with other systems). As another example, a component can be a device having specific functions provided by mechanical parts operated by an electrical or electronic circuitry system, which is operated by a software or firmware application executed by a processor. In this case, the processor can be internal or external to the device and can execute at least a portion of the software or firmware application. As yet another example, a component can be a device that provides specific functions through electronic components rather than mechanical parts, wherein the electronic components can include a processor or other means for executing software or firmware that at least partially gives the electronic components functionality. In one aspect, a component can be emulated, for example, via a virtual machine within a cloud computing system.

[0087] Furthermore, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless otherwise specified or clear from the context, "X adopts A or B" is intended to mean any natural inclusive substitution. That is, if X adopts A; X adopts B; or X adopts both A and B, then "X adopts A or B" is satisfied in any of the foregoing cases. Additionally, unless otherwise specified or clear from the context to be directed to the singular form, the articles "a" and "an" used in this specification and figures should generally be construed as meaning "one or more". As used herein, the terms "example" and / or "exemplary" are used to mean serving as an example, instance, or illustration, and are intended to be non-limiting. For the avoidance of doubt, the subject matter disclosed herein is not limited to such examples. Furthermore, any aspect or design described herein as "example" and / or "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects or designs, nor does it exclude equivalent exemplary structures and techniques known to those skilled in the art.

[0088] As used herein, the term "processor" can refer substantially to any computing processing unit or device, including but not limited to a single-core processor; a single processor with software multithreading capabilities; a multi-core processor; a multi-core processor with software multithreading capabilities; a multi-core processor with hardware multithreading technology; a parallel platform; and a parallel platform with distributed shared memory. Additionally, a processor can refer to an integrated circuit, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic controller (PLC), a complex programmable logic device (CPLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. Furthermore, processors can utilize nanoscale architectures (such as, but not limited to, molecular and quantum dot-based transistors, switches, and gates) to optimize space usage or enhance the performance of physical equipment. Processors can also be implemented as a combination of computing processing units. In this disclosure, terms such as "storage," "storage device," "data storage," "data storage apparatus," "database," and substantially any other information storage component associated with the operation and function of a component are used to refer to a "memory component," an entity embodied in "memory," or a component that includes memory. It should be understood that the memory and / or memory components described herein may be volatile or non-volatile memory, or may include both volatile and non-volatile memory. By way of illustration and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), flash memory, or non-volatile random access memory (RAM) (e.g., ferroelectric RAM (FeRAM)). For example, volatile memory may include RAM that can act as external cache memory. By way of illustration and not limitation, RAM can be provided in a variety of forms, such as synchronous RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), direct Rambus RAM (DRRAM), direct Rambus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM). Furthermore, the memory components disclosed in the systems or computer-implemented methods described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0089] The foregoing description includes only examples of systems and computer-implemented methods. Of course, it is impossible to describe every conceivable combination of components or computer-implemented methods for the purposes of describing this disclosure, but those skilled in the art will recognize that many other combinations and substitutions are possible. Furthermore, regarding the extent to which the terms “comprising,” “having,” “possessing,” etc., are used in the detailed description, claims, appendices, and drawings, such terms are intended to be inclusive in a manner similar to the term “comprising,” as interpreted when “comprising” is used as a transitional word in the claims. Various embodiments have been described for illustrative purposes, but these descriptions are not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein has been chosen to best explain the principles of the embodiments, the practical application or improvement of technology superior to that found in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A system for training a medical machine learning model, the system comprising: Memory, which stores computer-executable components; and A processor that executes computer-executable components stored in the memory, wherein the computer-executable components include: A training component trains a machine learning model on training data including medical images using an equalization loss function to perform a clinical inference task, thereby producing a trained version of the machine learning model. The equalization loss function drives the equalization performance of the machine learning model across different subgroups represented in the training data, and the trained version provides a defined level of equalization performance across the different subgroups.

2. The system for training a medical machine learning model according to claim 1, wherein, The distinct subgroups are labeled in the training data, and the equalization loss function is configured to use the labels to identify the distinct subgroups.

3. The system for training a medical machine learning model according to claim 1, wherein, The equilibrium loss function penalizes the performance variation of the machine learning model across the different subgroups.

4. The system for training a medical machine learning model according to claim 1, wherein, The equilibrium loss function employs different loss weights for at least two subgroups within the different subgroups.

5. The system for training a medical machine learning model according to claim 4, wherein, The different loss weights are based on the relative number of training data samples in the respective subgroups included in the different subgroups.

6. The system for training a medical machine learning model according to claim 1, wherein, The training component measures the performance of the machine learning model across the different subgroups based on the output of the equalization loss function, and wherein the computer-executable component further includes: A reporting component that generates a performance report that identifies the performance of the machine learning model across the different subgroups.

7. The system for training a medical machine learning model according to claim 1, wherein, The training component uses a supervised machine learning process that includes a training phase and a validation phase to train the machine learning model, wherein the training component uses the balanced loss function during the training phase and the validation phase.

8. The system for training a medical machine learning model according to claim 1, wherein, The training component uses a supervised machine learning process, including a training phase and a validation phase, to train the machine learning model, and wherein the computer-executable component further includes: A training data management component separates the training data into a training set and a validation set, and further identifies the different subgroups within the validation set, wherein, in association with minimizing the loss across the different subgroups, the training component uses the equalization loss function on the validation set during the validation phase.

9. A method for training a medical machine learning model, the method comprising: A machine learning model is trained on training data including medical images by a system including a processor using an equalization loss function to perform a clinical inference task, the equalization loss function driving the equalization performance of the machine learning model across different subgroups represented in the training data; as well as The system generates a trained version of the machine learning model based on the training, the trained version providing a defined, balanced performance level across the different subgroups.

10. The method for training a medical machine learning model according to claim 9, wherein, The distinct subgroups are labeled in the training data, and the equalization loss function is configured to use the labels to identify the distinct subgroups.

11. The method for training a medical machine learning model according to claim 9, wherein, The equilibrium loss function penalizes the performance variation of the machine learning model across the different subgroups.

12. The method for training a medical machine learning model according to claim 9, wherein, The equilibrium loss function employs different loss weights for at least two subgroups within the different subgroups.

13. The method for training a medical machine learning model according to claim 12, wherein, The different loss weights are based on the relative number of training data samples included in the respective subgroups of the different subgroups.

14. The method for training a medical machine learning model according to claim 9, the method further comprising: The system measures the performance of the machine learning model across the different subgroups based on the output of the equilibrium loss function; as well as The system generates a performance report that identifies the performance of the machine learning model across the different subgroups.

15. The method for training a medical machine learning model according to claim 9, wherein, The training includes a training phase and a validation phase, and the training includes employing the equalization loss function during the training phase and the validation phase.

16. The method for training a medical machine learning model according to claim 9, wherein, The training includes a training phase and a validation phase, and the method further includes: The system groups the training data into a training set and a validation set; and The system groups the validation set into subsets corresponding to the different subgroups, and the training includes applying the balanced loss function on the validation set during the validation phase, in association with minimizing the loss across the different subgroups.

17. A non-transitory machine-readable storage medium comprising executable instructions for training a medical machine learning model, the executable instructions facilitating the execution of operations when executed by a processor, the operations comprising: A machine learning model is trained on training data including medical images using an equalization loss function to perform a clinical inference task. The equalization loss function drives the equalization performance of the machine learning model across different subgroups represented in the training data. as well as The machine learning model is generated based on the training, and the trained version provides a defined, balanced performance level across the different subgroups.

18. The non-transitory machine-readable storage medium comprising executable instructions for training a medical machine learning model according to claim 17, wherein, The different subgroups are labeled in the training data, and the equalization loss function is configured to use the labels to identify the different subgroups and to penalize changes in the performance of the machine learning model across the different subgroups.

19. The non-transitory machine-readable storage medium comprising executable instructions for training a medical machine learning model according to claim 17, wherein, The equilibrium loss function employs different loss weights for at least two subgroups within the different subgroups.

20. The non-transitory machine-readable storage medium comprising executable instructions for training a medical machine learning model according to claim 19, wherein, The different loss weights are based on the relative number of training data samples in the respective subgroups included in the different subgroups.

Images