Judy Wawira Gichoya

Karim Lekadir, Aasa Feragen, Abdul Joseph Fofanah et al. · eth-zurich

Despite major advances in artificial intelligence (AI) for medicine and healthcare, the deployment and adoption of AI technologies remain limited in real-world clinical practice. In recent years, concerns have been raised about the technical, clinical, ethical and legal risks associated with medical AI. To increase real world adoption, it is essential that medical AI tools are trusted and accepted by patients, clinicians, health organisations and authorities. This work describes the FUTURE-AI guideline as the first international consensus framework for guiding the development and deployment of trustworthy AI tools in healthcare. The FUTURE-AI consortium was founded in 2021 and currently comprises 118 inter-disciplinary experts from 51 countries representing all continents, including AI scientists, clinicians, ethicists, and social scientists. Over a two-year period, the consortium defined guiding principles and best practices for trustworthy AI through an iterative process comprising an in-depth literature review, a modified Delphi survey, and online consensus meetings. The FUTURE-AI framework was established based on 6 guiding principles for trustworthy AI in healthcare, i.e. Fairness, Universality, Traceability, Usability, Robustness and Explainability. Through consensus, a set of 28 best practices were defined, addressing technical, clinical, legal and socio-ethical dimensions. The recommendations cover the entire lifecycle of medical AI, from design, development and validation to regulation, deployment, and monitoring. FUTURE-AI is a risk-informed, assumption-free guideline which provides a structured approach for constructing medical AI tools that will be trusted, deployed and adopted in real-world practice. Researchers are encouraged to take the recommendations into account in proof-of-concept stages to facilitate future translation towards clinical practice of medical AI.

Saptarshi Purkayastha, Rohan Isaac, Sharon Anthony et al.

Radiology AI models have made significant progress in near-human performance or surpassing it. However, AI model's partnership with human radiologist remains an unexplored challenge due to the lack of health information standards, contextual and workflow differences, and data labeling variations. To overcome these challenges, we integrated an AI model service that uses DICOM standard SR annotations into the OHIF viewer in the open-source LibreHealth Radiology Information Systems (RIS). In this paper, we describe the novel Human-AI partnership capabilities of the platform, including few-shot learning and swarm learning approaches to retrain the AI models continuously. Building on the concept of machine teaching, we developed an active learning strategy within the RIS, so that the human radiologist can enable/disable AI annotations as well as "fix"/relabel the AI annotations. These annotations are then used to retrain the models. This helps establish a partnership between the radiologist user and a user-specific AI model. The weights of these user-specific models are then finally shared between multiple models in a swarm learning approach.

Xiaoyuan Guo, Jiali Duan, Saptarshi Purkayastha et al.

Improving the retrieval relevance on noisy datasets is an emerging need for the curation of a large-scale clean dataset in the medical domain. While existing methods can be applied for class-wise retrieval (aka. inter-class), they cannot distinguish the granularity of likeness within the same class (aka. intra-class). The problem is exacerbated on medical external datasets, where noisy samples of the same class are treated equally during training. Our goal is to identify both intra/inter-class similarities for fine-grained retrieval. To achieve this, we propose an Outlier-Sensitive Content-based rAdiologhy Retrieval System (OSCARS), consisting of two steps. First, we train an outlier detector on a clean internal dataset in an unsupervised manner. Then we use the trained detector to generate the anomaly scores on the external dataset, whose distribution will be used to bin intra-class variations. Second, we propose a quadruplet (a, p, nintra, ninter) sampling strategy, where intra-class negatives nintra are sampled from bins of the same class other than the bin anchor a belongs to, while niner are randomly sampled from inter-classes. We suggest a weighted metric learning objective to balance the intra and inter-class feature learning. We experimented on two representative public radiography datasets. Experiments show the effectiveness of our approach. The training and evaluation code can be found in https://github.com/XiaoyuanGuo/oscars.

Bardia Khosravi, Frank Li, Theo Dapamede et al.

Chest X-rays (CXR) are essential for diagnosing a variety of conditions, but when used on new populations, model generalizability issues limit their efficacy. Generative AI, particularly denoising diffusion probabilistic models (DDPMs), offers a promising approach to generating synthetic images, enhancing dataset diversity. This study investigates the impact of synthetic data supplementation on the performance and generalizability of medical imaging research. The study employed DDPMs to create synthetic CXRs conditioned on demographic and pathological characteristics from the CheXpert dataset. These synthetic images were used to supplement training datasets for pathology classifiers, with the aim of improving their performance. The evaluation involved three datasets (CheXpert, MIMIC-CXR, and Emory Chest X-ray) and various experiments, including supplementing real data with synthetic data, training with purely synthetic data, and mixing synthetic data with external datasets. Performance was assessed using the area under the receiver operating curve (AUROC). Adding synthetic data to real datasets resulted in a notable increase in AUROC values (up to 0.02 in internal and external test sets with 1000% supplementation, p-value less than 0.01 in all instances). When classifiers were trained exclusively on synthetic data, they achieved performance levels comparable to those trained on real data with 200%-300% data supplementation. The combination of real and synthetic data from different sources demonstrated enhanced model generalizability, increasing model AUROC from 0.76 to 0.80 on the internal test set (p-value less than 0.01). In conclusion, synthetic data supplementation significantly improves the performance and generalizability of pathology classifiers in medical imaging.

David A. Clunie, Adam Flanders, Adam Taylor et al.

This report addresses the technical aspects of de-identification of medical images of human subjects and biospecimens, such that re-identification risk of ethical, moral, and legal concern is sufficiently reduced to allow unrestricted public sharing for any purpose, regardless of the jurisdiction of the source and distribution sites. All medical images, regardless of the mode of acquisition, are considered, though the primary emphasis is on those with accompanying data elements, especially those encoded in formats in which the data elements are embedded, particularly Digital Imaging and Communications in Medicine (DICOM). These images include image-like objects such as Segmentations, Parametric Maps, and Radiotherapy (RT) Dose objects. The scope also includes related non-image objects, such as RT Structure Sets, Plans and Dose Volume Histograms, Structured Reports, and Presentation States. Only de-identification of publicly released data is considered, and alternative approaches to privacy preservation, such as federated learning for artificial intelligence (AI) model development, are out of scope, as are issues of privacy leakage from AI model sharing. Only technical issues of public sharing are addressed.

Xiaoyuan Guo, Jiali Duan, C. -C. Jay Kuo et al.

Language modality within the vision language pretraining framework is innately discretized, endowing each word in the language vocabulary a semantic meaning. In contrast, visual modality is inherently continuous and high-dimensional, which potentially prohibits the alignment as well as fusion between vision and language modalities. We therefore propose to "discretize" the visual representation by joint learning a codebook that imbues each visual token a semantic. We then utilize these discretized visual semantics as self-supervised ground-truths for building our Masked Image Modeling objective, a counterpart of Masked Language Modeling which proves successful for language models. To optimize the codebook, we extend the formulation of VQ-VAE which gives a theoretic guarantee. Experiments validate the effectiveness of our approach across common vision-language benchmarks.

Ryan Wang, Li-Ching Chen, Lama Moukheiber et al.

Chronic obstructive pulmonary disease (COPD) is one of the most common chronic illnesses in the world and the third leading cause of mortality worldwide. It is often underdiagnosed or not diagnosed until later in the disease course. Spirometry tests are the gold standard for diagnosing COPD but can be difficult to obtain, especially in resource-poor countries. Chest X-rays (CXRs), however, are readily available and may serve as a screening tool to identify patients with COPD who should undergo further testing. Currently, no research applies deep learning (DL) algorithms that use large multi-site and multi-modal data to detect COPD patients and evaluate fairness across demographic groups. We use three CXR datasets in our study, CheXpert to pre-train models, MIMIC-CXR to develop, and Emory-CXR to validate our models. The CXRs from patients in the early stage of COPD and not on mechanical ventilation are selected for model training and validation. We visualize the Grad-CAM heatmaps of the true positive cases on the base model for both MIMIC-CXR and Emory-CXR test datasets. We further propose two fusion schemes, (1) model-level fusion, including bagging and stacking methods using MIMIC-CXR, and (2) data-level fusion, including multi-site data using MIMIC-CXR and Emory-CXR, and multi-modal using MIMIC-CXRs and MIMIC-IV EHR, to improve the overall model performance. Fairness analysis is performed to evaluate if the fusion schemes have a discrepancy in the performance among different demographic groups. The results demonstrate that DL models can detect COPD using CXRs, which can facilitate early screening, especially in low-resource regions where CXRs are more accessible than spirometry. The multi-site data fusion scheme could improve the model generalizability on the Emory-CXR test data. Further studies on using CXR or other modalities to predict COPD ought to be in future work.

Shuja Khalid, Francisco Matos, Ayman Abunimer et al.

30-day hospital readmission is a long standing medical problem that affects patients' morbidity and mortality and costs billions of dollars annually. Recently, machine learning models have been created to predict risk of inpatient readmission for patients with specific diseases, however no model exists to predict this risk across all patients. We developed a bi-directional Long Short Term Memory (LSTM) Network that is able to use readily available insurance data (inpatient visits, outpatient visits, and drug prescriptions) to predict 30 day re-admission for any admitted patient, regardless of reason. The top-performing model achieved an ROC AUC of 0.763 (0.011) when using historical, inpatient, and post-discharge data. The LSTM model significantly outperformed a baseline random forest classifier, indicating that understanding the sequence of events is important for model prediction. Incorporation of 30-days of historical data also significantly improved model performance compared to inpatient data alone, indicating that a patients clinical history prior to admission, including outpatient visits and pharmacy data is a strong contributor to readmission. Our results demonstrate that a machine learning model is able to predict risk of inpatient readmission with reasonable accuracy for all patients using structured insurance billing data. Because billing data or equivalent surrogates can be extracted from sites, such a model could be deployed to identify patients at risk for readmission before they are discharged, or to assign more robust follow up (closer follow up, home health, mailed medications) to at-risk patients after discharge.

Carolina A. M. Heming, Mohamed Abdalla, Shahram Mohanna et al.

Clinical AI model reporting cards should be expanded to incorporate a broad bias reporting of both social and non-social factors. Non-social factors consider the role of other factors, such as disease dependent, anatomic, or instrument factors on AI model bias, which are essential to ensure safe deployment.

Ananth Reddy Bhimireddy, John Lee Burns, Saptarshi Purkayastha et al.

Deep learning approaches applied to medical imaging have reached near-human or better-than-human performance on many diagnostic tasks. For instance, the CheXpert competition on detecting pathologies in chest x-rays has shown excellent multi-class classification performance. However, training and validating deep learning models require extensive collections of images and still produce false inferences, as identified by a human-in-the-loop. In this paper, we introduce a practical approach to improve the predictions of a pre-trained model through Few-Shot Learning (FSL). After training and validating a model, a small number of false inference images are collected to retrain the model using \textbf{\textit{Image Triplets}} - a false positive or false negative, a true positive, and a true negative. The retrained FSL model produces considerable gains in performance with only a few epochs and few images. In addition, FSL opens rapid retraining opportunities for human-in-the-loop systems, where a radiologist can relabel false inferences, and the model can be quickly retrained. We compare our retrained model performance with existing FSL approaches in medical imaging that train and evaluate models at once.

Frank Li, Theo Dapamede, Mohammadreza Chavoshi et al.

Foundation models (FMs) promise to generalize medical imaging, but their effectiveness varies. It remains unclear how pre-training domain (medical vs. general), paradigm (e.g., text-guided), and architecture influence embedding quality, hindering the selection of optimal encoders for specific radiology tasks. To address this, we evaluate vision encoders from eight medical and general-domain FMs for chest X-ray analysis. We benchmark classification (pneumothorax, cardiomegaly) and segmentation (pneumothorax, cardiac boundary) using linear probing and fine-tuning. Our results show that domain-specific pre-training provides a significant advantage; medical FMs consistently outperformed general-domain models in linear probing, establishing superior initial feature quality. However, feature utility is highly task-dependent. Pre-trained embeddings were strong for global classification and segmenting salient anatomy (e.g., heart). In contrast, for segmenting complex, subtle pathologies (e.g., pneumothorax), all FMs performed poorly without significant fine-tuning, revealing a critical gap in localizing subtle disease. Subgroup analysis showed FMs use confounding shortcuts (e.g., chest tubes for pneumothorax) for classification, a strategy that fails for precise segmentation. We also found that expensive text-image alignment is not a prerequisite; image-only (RAD-DINO) and label-supervised (Ark+) FMs were among top performers. Notably, a supervised, end-to-end baseline remained highly competitive, matching or exceeding the best FMs on segmentation tasks. These findings show that while medical pre-training is beneficial, architectural choices (e.g., multi-scale) are critical, and pre-trained features are not universally effective, especially for complex localization tasks where supervised models remain a strong alternative.

Thiago Santos, Harish Kamath, Christopher R. McAdams et al.

Automated classification of cancer pathology reports can extract information from unstructured reports and categorize each report into structured diagnosis and severity categories. Thus, such system can reduce the burden for populating tumor registries, help registration for clinical trial as well as developing large dataset for deep learning model development using true pathologic ground truth. However, the content of breast pathology reports can be difficult for categorize due to the high linguistic variability in content and wide variety of potential diagnoses >50. Existing NLP models are primarily focused on developing classifier for primary breast cancer types (e.g. IDC, DCIS, ILC) and tumor characteristics, and ignore the rare diagnosis of cancer subtypes. We then developed a hierarchical hybrid transformer-based pipeline (59 labels) - Hierarchical Classification System for Breast Cancer Specimen Report (HCSBC), which utilizes the potential of the transformer context-preserving NLP technique and compared our model to several state of the art ML and DL models. We trained the model on the EUH data and evaluated our model's performance on two external datasets - MGH and Mayo Clinic. We publicly release the code and a live application under Huggingface spaces repository

Xiaoyuan Guo, Judy Wawira Gichoya, Saptarshi Purkayastha et al.

Detecting out-of-distribution (OOD) samples in medical imaging plays an important role for downstream medical diagnosis. However, existing OOD detectors are demonstrated on natural images composed of inter-classes and have difficulty generalizing to medical images. The key issue is the granularity of OOD data in the medical domain, where intra-class OOD samples are predominant. We focus on the generalizability of OOD detection for medical images and propose a self-supervised Cascade Variational autoencoder-based Anomaly Detector (CVAD). We use a variational autoencoders' cascade architecture, which combines latent representation at multiple scales, before being fed to a discriminator to distinguish the OOD data from the in-distribution (ID) data. Finally, both the reconstruction error and the OOD probability predicted by the binary discriminator are used to determine the anomalies. We compare the performance with the state-of-the-art deep learning models to demonstrate our model's efficacy on various open-access medical imaging datasets for both intra- and inter-class OOD. Further extensive results on datasets including common natural datasets show our model's effectiveness and generalizability. The code is available at https://github.com/XiaoyuanGuo/CVAD.

Amelia Jiménez-Sánchez, Natalia-Rozalia Avlona, Sarah de Boer et al.

Datasets play a critical role in medical imaging research, yet issues such as label quality, shortcuts, and metadata are often overlooked. This lack of attention may harm the generalizability of algorithms and, consequently, negatively impact patient outcomes. While existing medical imaging literature reviews mostly focus on machine learning (ML) methods, with only a few focusing on datasets for specific applications, these reviews remain static -- they are published once and not updated thereafter. This fails to account for emerging evidence, such as biases, shortcuts, and additional annotations that other researchers may contribute after the dataset is published. We refer to these newly discovered findings of datasets as research artifacts. To address this gap, we propose a living review that continuously tracks public datasets and their associated research artifacts across multiple medical imaging applications. Our approach includes a framework for the living review to monitor data documentation artifacts, and an SQL database to visualize the citation relationships between research artifact and dataset. Lastly, we discuss key considerations for creating medical imaging datasets, review best practices for data annotation, discuss the significance of shortcuts and demographic diversity, and emphasize the importance of managing datasets throughout their entire lifecycle. Our demo is publicly available at http://inthepicture.itu.dk/.

Yuzhe Yang, Haoran Zhang, Judy W Gichoya et al.

As artificial intelligence (AI) rapidly approaches human-level performance in medical imaging, it is crucial that it does not exacerbate or propagate healthcare disparities. Prior research has established AI's capacity to infer demographic data from chest X-rays, leading to a key concern: do models using demographic shortcuts have unfair predictions across subpopulations? In this study, we conduct a thorough investigation into the extent to which medical AI utilizes demographic encodings, focusing on potential fairness discrepancies within both in-distribution training sets and external test sets. Our analysis covers three key medical imaging disciplines: radiology, dermatology, and ophthalmology, and incorporates data from six global chest X-ray datasets. We confirm that medical imaging AI leverages demographic shortcuts in disease classification. While correcting shortcuts algorithmically effectively addresses fairness gaps to create "locally optimal" models within the original data distribution, this optimality is not true in new test settings. Surprisingly, we find that models with less encoding of demographic attributes are often most "globally optimal", exhibiting better fairness during model evaluation in new test environments. Our work establishes best practices for medical imaging models which maintain their performance and fairness in deployments beyond their initial training contexts, underscoring critical considerations for AI clinical deployments across populations and sites.

Jacki O'Neill, Vukosi Marivate, Barbara Glover et al.

This white paper is the output of a multidisciplinary workshop in Nairobi (Nov 2023). Led by a cross-organisational team including Microsoft Research, NEPAD, Lelapa AI, and University of Oxford. The workshop brought together diverse thought-leaders from various sectors and backgrounds to discuss the implications of Generative AI for the future of work in Africa. Discussions centred around four key themes: Macroeconomic Impacts; Jobs, Skills and Labour Markets; Workers' Perspectives and Africa-Centris AI Platforms. The white paper provides an overview of the current state and trends of generative AI and its applications in different domains, as well as the challenges and risks associated with its adoption and regulation. It represents a diverse set of perspectives to create a set of insights and recommendations which aim to encourage debate and collaborative action towards creating a dignified future of work for everyone across Africa.

Lasse Hyldig Hansen, Nikolaj Andersen, Jack Gallifant et al.

Background Advancements in Large Language Models (LLMs) hold transformative potential in healthcare, however, recent work has raised concern about the tendency of these models to produce outputs that display racial or gender biases. Although training data is a likely source of such biases, exploration of disease and demographic associations in text data at scale has been limited. Methods We conducted a large-scale textual analysis using a dataset comprising diverse web sources, including Arxiv, Wikipedia, and Common Crawl. The study analyzed the context in which various diseases are discussed alongside markers of race and gender. Given that LLMs are pre-trained on similar datasets, this approach allowed us to examine the potential biases that LLMs may learn and internalize. We compared these findings with actual demographic disease prevalence as well as GPT-4 outputs in order to evaluate the extent of bias representation. Results Our findings indicate that demographic terms are disproportionately associated with specific disease concepts in online texts. gender terms are prominently associated with disease concepts, while racial terms are much less frequently associated. We find widespread disparities in the associations of specific racial and gender terms with the 18 diseases analyzed. Most prominently, we see an overall significant overrepresentation of Black race mentions in comparison to population proportions. Conclusions Our results highlight the need for critical examination and transparent reporting of biases in LLM pretraining datasets. Our study suggests the need to develop mitigation strategies to counteract the influence of biased training data in LLMs, particularly in sensitive domains such as healthcare.

Jungkyu Park, Jan Witowski, Yanqi Xu et al.

Although digital breast tomosynthesis (DBT) improves diagnostic performance over full-field digital mammography (FFDM), false-positive recalls remain a concern in breast cancer screening. We developed a multi-modal artificial intelligence system integrating FFDM, synthetic mammography, and DBT to provide breast-level predictions and bounding-box localizations of suspicious findings. Our AI system, trained on approximately 500,000 mammography exams, achieved 0.945 AUROC on an internal test set. It demonstrated capacity to reduce recalls by 31.7% and radiologist workload by 43.8% while maintaining 100% sensitivity, underscoring its potential to improve clinical workflows. External validation confirmed strong generalizability, reducing the gap to a perfect AUROC by 35.31%-69.14% relative to strong baselines. In prospective deployment across 18 sites, the system reduced recall rates for low-risk cases. An improved version, trained on over 750,000 exams with additional labels, further reduced the gap by 18.86%-56.62% across large external datasets. Overall, these results underscore the importance of utilizing all available imaging modalities, demonstrate the potential for clinical impact, and indicate feasibility of further reduction of the test error with increased training set when using large-capacity neural networks.

Theodorus Dapamede, Aisha Urooj, Vedant Joshi et al.

Women are underdiagnosed and undertreated for cardiovascular disease. Automatic quantification of breast arterial calcification on screening mammography can identify women at risk for cardiovascular disease and enable earlier treatment and management of disease. In this retrospective study of 116,135 women from two healthcare systems, a transformer-based neural network quantified BAC severity (no BAC, mild, moderate, and severe) on screening mammograms. Outcomes included major adverse cardiovascular events (MACE) and all-cause mortality. BAC severity was independently associated with MACE after adjusting for cardiovascular risk factors, with increasing hazard ratios from mild (HR 1.18-1.22), moderate (HR 1.38-1.47), to severe BAC (HR 2.03-2.22) across datasets (all p<0.001). This association remained significant across all age groups, with even mild BAC indicating increased risk in women under 50. BAC remained an independent predictor when analyzed alongside ASCVD risk scores, showing significant associations with myocardial infarction, stroke, heart failure, and mortality (all p<0.005). Automated BAC quantification enables opportunistic cardiovascular risk assessment during routine mammography without additional radiation or cost. This approach provides value beyond traditional risk factors, particularly in younger women, offering potential for early CVD risk stratification in the millions of women undergoing annual mammography.

Shantanu Ghosh, Vedant Parthesh Joshi, Rayan Syed et al.

Breast cancer is one of the leading causes of death among women worldwide. We introduce Mammo-FM, the first foundation model specifically for mammography, pretrained on the largest and most diverse dataset to date - 140,677 patients (821,326 mammograms) across four U.S. institutions. Mammo-FM provides a unified foundation for core clinical tasks in breast imaging, including cancer diagnosis, pathology localization, structured report generation, and cancer risk prognosis within a single framework. Its alignment between images and text enables both visual and textual interpretability, improving transparency and clinical auditability, which are essential for real-world adoption. We rigorously evaluate Mammo-FM across diagnosis, prognosis, and report-generation tasks in in- and out-of-distribution datasets. Despite operating on native-resolution mammograms and using only one-third of the parameters of state-of-the-art generalist FMs, Mammo-FM consistently outperforms them across multiple public and private benchmarks. These results highlight the efficiency and value of domain-specific foundation models designed around the full spectrum of tasks within a clinical domain and emphasize the importance of rigorous, domain-aligned evaluation.

Emily Alsentzer, Marie-Laure Charpignon, Bill Chen et al.

The 6th Annual Conference on Health, Inference, and Learning (CHIL 2025), hosted by the Association for Health Learning and Inference (AHLI), was held in person on June 25-27, 2025, at the University of California, Berkeley, in Berkeley, California, USA. As part of this year's program, we hosted Research Roundtables to catalyze collaborative, small-group dialogue around critical, timely topics at the intersection of machine learning and healthcare. Each roundtable was moderated by a team of senior and junior chairs who fostered open exchange, intellectual curiosity, and inclusive engagement. The sessions emphasized rigorous discussion of key challenges, exploration of emerging opportunities, and collective ideation toward actionable directions in the field. In total, eight roundtables were held by 19 roundtable chairs on topics of "Explainability, Interpretability, and Transparency," "Uncertainty, Bias, and Fairness," "Causality," "Domain Adaptation," "Foundation Models," "Learning from Small Medical Data," "Multimodal Methods," and "Scalable, Translational Healthcare Solutions."

Frank Li, Hari Trivedi, Bardia Khosravi et al.

Foundation models, trained on vast amounts of data using self-supervised techniques, have emerged as a promising frontier for advancing artificial intelligence (AI) applications in medicine. This study evaluates three different vision-language foundation models (RAD-DINO, CheXagent, and BiomedCLIP) on their ability to capture fine-grained imaging features for radiology tasks. The models were assessed across classification, segmentation, and regression tasks for pneumothorax and cardiomegaly on chest radiographs. Self-supervised RAD-DINO consistently excelled in segmentation tasks, while text-supervised CheXagent demonstrated superior classification performance. BiomedCLIP showed inconsistent performance across tasks. A custom segmentation model that integrates global and local features substantially improved performance for all foundation models, particularly for challenging pneumothorax segmentation. The findings highlight that pre-training methodology significantly influences model performance on specific downstream tasks. For fine-grained segmentation tasks, models trained without text supervision performed better, while text-supervised models offered advantages in classification and interpretability. These insights provide guidance for selecting foundation models based on specific clinical applications in radiology.

Brandon Price, Jason Adleberg, Kaesha Thomas et al.

The Emory Knee Radiograph (MRKR) dataset is a large, demographically diverse collection of 503,261 knee radiographs from 83,011 patients, 40% of which are African American. This dataset provides imaging data in DICOM format along with detailed clinical information, including patient-reported pain scores, diagnostic codes, and procedural codes, which are not commonly available in similar datasets. The MRKR dataset also features imaging metadata such as image laterality, view type, and presence of hardware, enhancing its value for research and model development. MRKR addresses significant gaps in existing datasets by offering a more representative sample for studying osteoarthritis and related outcomes, particularly among minority populations, thereby providing a valuable resource for clinicians and researchers.

Linglin Zhang, Beatrice Brown-Mulry, Vineela Nalla et al.

Although deep learning models for abnormality classification can perform well in screening mammography, the demographic, imaging, and clinical characteristics associated with increased risk of model failure remain unclear. This retrospective study uses the Emory BrEast Imaging Dataset(EMBED) containing mammograms from 115931 patients imaged at Emory Healthcare between 2013-2020, with BI-RADS assessment, region of interest coordinates for abnormalities, imaging features, pathologic outcomes, and patient demographics. Multiple deep learning models were trained to distinguish between abnormal tissue patches and randomly selected normal tissue patches from screening mammograms. We assessed model performance by subgroups defined by age, race, pathologic outcome, tissue density, and imaging characteristics and investigated their associations with false negatives (FN) and false positives (FP). We also performed multivariate logistic regression to control for confounding between subgroups. The top-performing model, ResNet152V2, achieved accuracy of 92.6%(95%CI=92.0-93.2%), and AUC 0.975(95%CI=0.972-0.978). Before controlling for confounding, nearly all subgroups showed statistically significant differences in model performance. However, after controlling for confounding, we found lower FN risk associates with Other race(RR=0.828;p=.050), biopsy-proven benign lesions(RR=0.927;p=.011), and mass(RR=0.921;p=.010) or asymmetry(RR=0.854;p=.040); higher FN risk associates with architectural distortion (RR=1.037;p<.001). Higher FP risk associates to BI-RADS density C(RR=1.891;p<.001) and D(RR=2.486;p<.001). Our results demonstrate subgroup analysis is important in mammogram classifier performance evaluation, and controlling for confounding between subgroups elucidates the true associations between variables and model failure. These results can help guide developing future breast cancer detection models.

Jiwoong J. Jeong, Brianna L. Vey, Ananth Reddy et al.

Developing and validating artificial intelligence models in medical imaging requires datasets that are large, granular, and diverse. To date, the majority of publicly available breast imaging datasets lack in one or more of these areas. Models trained on these data may therefore underperform on patient populations or pathologies that have not previously been encountered. The EMory BrEast imaging Dataset (EMBED) addresses these gaps by providing 3650,000 2D and DBT screening and diagnostic mammograms for 116,000 women divided equally between White and African American patients. The dataset also contains 40,000 annotated lesions linked to structured imaging descriptors and 61 ground truth pathologic outcomes grouped into six severity classes. Our goal is to share this dataset with research partners to aid in development and validation of breast AI models that will serve all patients fairly and help decrease bias in medical AI.

Timothy L. Kline, Felipe Kitamura, Ian Pan et al.

With the recent advances in A.I. methodologies and their application to medical imaging, there has been an explosion of related research programs utilizing these techniques to produce state-of-the-art classification performance. Ultimately, these research programs culminate in submission of their work for consideration in peer reviewed journals. To date, the criteria for acceptance vs. rejection is often subjective; however, reproducible science requires reproducible review. The Machine Learning Education Sub-Committee of SIIM has identified a knowledge gap and a serious need to establish guidelines for reviewing these studies. Although there have been several recent papers with this goal, this present work is written from the machine learning practitioners standpoint. In this series, the committee will address the best practices to be followed in an A.I.-based study and present the required sections in terms of examples and discussion of what should be included to make the studies cohesive, reproducible, accurate, and self-contained. This first entry in the series focuses on the task of image classification. Elements such as dataset curation, data pre-processing steps, defining an appropriate reference standard, data partitioning, model architecture and training are discussed. The sections are presented as they would be detailed in a typical manuscript, with content describing the necessary information that should be included to make sure the study is of sufficient quality to be considered for publication. The goal of this series is to provide resources to not only help improve the review process for A.I.-based medical imaging papers, but to facilitate a standard for the information that is presented within all components of the research study. We hope to provide quantitative metrics in what otherwise may be a qualitative review process.

Xiaoyuan Guo, Judy Wawira Gichoya, Hari Trivedi et al.

To curate a high-quality dataset, identifying data variance between the internal and external sources is a fundamental and crucial step. However, methods to detect shift or variance in data have not been significantly researched. Challenges to this are the lack of effective approaches to learn dense representation of a dataset and difficulties of sharing private data across medical institutions. To overcome the problems, we propose a unified pipeline called MedShift to detect the top-level shift samples and thus facilitate the medical curation. Given an internal dataset A as the base source, we first train anomaly detectors for each class of dataset A to learn internal distributions in an unsupervised way. Second, without exchanging data across sources, we run the trained anomaly detectors on an external dataset B for each class. The data samples with high anomaly scores are identified as shift data. To quantify the shiftness of the external dataset, we cluster B's data into groups class-wise based on the obtained scores. We then train a multi-class classifier on A and measure the shiftness with the classifier's performance variance on B by gradually dropping the group with the largest anomaly score for each class. Additionally, we adapt a dataset quality metric to help inspect the distribution differences for multiple medical sources. We verify the efficacy of MedShift with musculoskeletal radiographs (MURA) and chest X-rays datasets from more than one external source. Experiments show our proposed shift data detection pipeline can be beneficial for medical centers to curate high-quality datasets more efficiently. An interface introduction video to visualize our results is available at https://youtu.be/V3BF0P1sxQE.

Ramon Correa, Jiwoong Jason Jeong, Bhavik Patel et al.

The use of artificial intelligence (AI) in healthcare has become a very active research area in the last few years. While significant progress has been made in image classification tasks, only a few AI methods are actually being deployed in hospitals. A major hurdle in actively using clinical AI models currently is the trustworthiness of these models. More often than not, these complex models are black boxes in which promising results are generated. However, when scrutinized, these models begin to reveal implicit biases during the decision making, such as detecting race and having bias towards ethnic groups and subpopulations. In our ongoing study, we develop a two-step adversarial debiasing approach with partial learning that can reduce the racial disparity while preserving the performance of the targeted task. The methodology has been evaluated on two independent medical image case-studies - chest X-ray and mammograms, and showed promises in bias reduction while preserving the targeted performance.

Xiaoyuan Guo, Judy Wawira Gichoya, Saptarshi Purkayastha et al.

Traditional anomaly detection methods focus on detecting inter-class variations while medical image novelty identification is inherently an intra-class detection problem. For example, a machine learning model trained with normal chest X-ray and common lung abnormalities, is expected to discover and flag idiopathic pulmonary fibrosis which a rare lung disease and unseen by the model during training. The nuances from intra-class variations and lack of relevant training data in medical image analysis pose great challenges for existing anomaly detection methods. To tackle the challenges, we propose a hybrid model - Transformation-based Embedding learning for Novelty Detection (TEND) which without any out-of-distribution training data, performs novelty identification by combining both autoencoder-based and classifier-based method. With a pre-trained autoencoder as image feature extractor, TEND learns to discriminate the feature embeddings of in-distribution data from the transformed counterparts as fake out-of-distribution inputs. To enhance the separation, a distance objective is optimized to enforce a margin between the two classes. Extensive experimental results on both natural image datasets and medical image datasets are presented and our method out-performs state-of-the-art approaches.

Imon Banerjee, Ananth Reddy Bhimireddy, John L. Burns et al.

Background: In medical imaging, prior studies have demonstrated disparate AI performance by race, yet there is no known correlation for race on medical imaging that would be obvious to the human expert interpreting the images. Methods: Using private and public datasets we evaluate: A) performance quantification of deep learning models to detect race from medical images, including the ability of these models to generalize to external environments and across multiple imaging modalities, B) assessment of possible confounding anatomic and phenotype population features, such as disease distribution and body habitus as predictors of race, and C) investigation into the underlying mechanism by which AI models can recognize race. Findings: Standard deep learning models can be trained to predict race from medical images with high performance across multiple imaging modalities. Our findings hold under external validation conditions, as well as when models are optimized to perform clinically motivated tasks. We demonstrate this detection is not due to trivial proxies or imaging-related surrogate covariates for race, such as underlying disease distribution. Finally, we show that performance persists over all anatomical regions and frequency spectrum of the images suggesting that mitigation efforts will be challenging and demand further study. Interpretation: We emphasize that model ability to predict self-reported race is itself not the issue of importance. However, our findings that AI can trivially predict self-reported race -- even from corrupted, cropped, and noised medical images -- in a setting where clinical experts cannot, creates an enormous risk for all model deployments in medical imaging: if an AI model secretly used its knowledge of self-reported race to misclassify all Black patients, radiologists would not be able to tell using the same data the model has access to.

Nazanin Mashhaditafreshi, Amara Tariq, Judy Wawira Gichoya et al.

Deep Convolutional Neural Networks (DCNNs) have attracted extensive attention and been applied in many areas, including medical image analysis and clinical diagnosis. One major challenge is to conceive a DCNN model with remarkable performance on both internal and external data. We demonstrate that DCNNs may not generalize to new data, but increasing the quality and heterogeneity of the training data helps to improve the generalizibility factor. We use InceptionResNetV2 and DenseNet121 architectures to predict the risk of 5 common chest pathologies. The experiments were conducted on three publicly available databases: CheXpert, ChestX-ray14, and MIMIC Chest Xray JPG. The results show the internal performance of each of the 5 pathologies outperformed external performance on both of the models. Moreover, our strategy of exposing the models to a mix of different datasets during the training phase helps to improve model performance on the external dataset.

Ningcheng Li, Guy Maresh, Maxwell Cretcher et al.

Healthcare data is increasing in size at an unprecedented speed with much attention on big data analysis and Artificial Intelligence application for quality assurance, clinical training, severity triaging, and decision support. Radiology is well-suited for innovation given its intrinsically paired linguistic and visual data. Previous attempts to unlock this information goldmine were encumbered by heterogeneity of human language, proprietary search algorithms, and lack of medicine-specific search performance matrices. We present a de novo process of developing a document-based, secure, efficient, and accurate search engine in the context of Radiology. We assess our implementation of the search engine with comparison to pre-existing manually collected clinical databases used previously for clinical research projects in addition to computational performance benchmarks and survey feedback. By leveraging efficient database architecture, search capability, and clinical thinking, radiologists are at the forefront of harnessing the power of healthcare data.

Imon Banerjee, Priyanshu Sinha, Saptarshi Purkayastha et al.

Purpose: Since the recent COVID-19 outbreak, there has been an avalanche of research papers applying deep learning based image processing to chest radiographs for detection of the disease. To test the performance of the two top models for CXR COVID-19 diagnosis on external datasets to assess model generalizability. Methods: In this paper, we present our argument regarding the efficiency and applicability of existing deep learning models for COVID-19 diagnosis. We provide results from two popular models - COVID-Net and CoroNet evaluated on three publicly available datasets and an additional institutional dataset collected from EMORY Hospital between January and May 2020, containing patients tested for COVID-19 infection using RT-PCR. Results: There is a large false positive rate (FPR) for COVID-Net on both ChexPert (55.3%) and MIMIC-CXR (23.4%) dataset. On the EMORY Dataset, COVID-Net has 61.4% sensitivity, 0.54 F1-score and 0.49 precision value. The FPR of the CoroNet model is significantly lower across all the datasets as compared to COVID-Net - EMORY(9.1%), ChexPert (1.3%), ChestX-ray14 (0.02%), MIMIC-CXR (0.06%). Conclusion: The models reported good to excellent performance on their internal datasets, however we observed from our testing that their performance dramatically worsened on external data. This is likely from several causes including overfitting models due to lack of appropriate control patients and ground truth labels. The fourth institutional dataset was labeled using RT-PCR, which could be positive without radiographic findings and vice versa. Therefore, a fusion model of both clinical and radiographic data may have better performance and generalization.

Pradeeban Kathiravelu, Puneet Sharma, Ashish Sharma et al.

Executing machine learning (ML) pipelines in real-time on radiology images is hard due to the limited computing resources in clinical environments and the lack of efficient data transfer capabilities to run them on research clusters. We propose Niffler, an integrated framework that enables the execution of ML pipelines at research clusters by efficiently querying and retrieving radiology images from the Picture Archiving and Communication Systems (PACS) of the hospitals. Niffler uses the Digital Imaging and Communications in Medicine (DICOM) protocol to fetch and store imaging data and provides metadata extraction capabilities and Application programming interfaces (APIs) to apply filters on the images. Niffler further enables the sharing of the outcomes from the ML pipelines in a de-identified manner. Niffler has been running stable for more than 19 months and has supported several research projects at the department. In this paper, we present its architecture and three of its use cases: an inferior vena cava (IVC) filter detection from the images in real-time, identification of scanner utilization, and scanner clock calibration. Evaluations on the Niffler prototype highlight its feasibility and efficiency in facilitating the ML pipelines on the images and metadata in real-time and retrospectively.

A. K. Bhavani Singh, Mounika Guntu, Ananth Reddy Bhimireddy et al.

In the United States, 25% or greater than 200 billion dollars of hospital spending accounts for administrative costs that involve services for medical coding and billing. With the increasing number of patient records, manual assignment of the codes performed is overwhelming, time-consuming and error-prone, causing billing errors. Natural language processing can automate the extraction of codes/labels from unstructured clinical notes, which can aid human coders to save time, increase productivity, and verify medical coding errors. Our objective is to identify appropriate diagnosis and procedure codes from clinical notes by performing multi-label classification. We used de-identified data of critical care patients from the MIMIC-III database and subset the data to select the ten (top-10) and fifty (top-50) most common diagnoses and procedures, which covers 47.45% and 74.12% of all admissions respectively. We implemented state-of-the-art Bidirectional Encoder Representations from Transformers (BERT) to fine-tune the language model on 80% of the data and validated on the remaining 20%. The model achieved an overall accuracy of 87.08%, an F1 score of 85.82%, and an AUC of 91.76% for top-10 codes. For the top-50 codes, our model achieved an overall accuracy of 93.76%, an F1 score of 92.24%, and AUC of 91%. When compared to previously published research, our model outperforms in predicting codes from the clinical text. We discuss approaches to generalize the knowledge discovery process of our MIMIC-BERT to other clinical notes. This can help human coders to save time, prevent backlogs, and additional costs due to coding errors.

Siddhartha Nuthakki, Sunil Neela, Judy W. Gichoya et al.

Coding diagnosis and procedures in medical records is a crucial process in the healthcare industry, which includes the creation of accurate billings, receiving reimbursements from payers, and creating standardized patient care records. In the United States, Billing and Insurance related activities cost around $471 billion in 2012 which constitutes about 25% of all the U.S hospital spending. In this paper, we report the performance of a natural language processing model that can map clinical notes to medical codes, and predict final diagnosis from unstructured entries of history of present illness, symptoms at the time of admission, etc. Previous studies have demonstrated that deep learning models perform better at such mapping when compared to conventional machine learning models. Therefore, we employed state-of-the-art deep learning method, ULMFiT on the largest emergency department clinical notes dataset MIMIC III which has 1.2M clinical notes to select for the top-10 and top-50 diagnosis and procedure codes. Our models were able to predict the top-10 diagnoses and procedures with 80.3% and 80.5% accuracy, whereas the top-50 ICD-9 codes of diagnosis and procedures are predicted with 70.7% and 63.9% accuracy. Prediction of diagnosis and procedures from unstructured clinical notes benefit human coders to save time, eliminate errors and minimize costs. With promising scores from our present model, the next step would be to deploy this on a small-scale real-world scenario and compare it with human coders as the gold standard. We believe that further research of this approach can create highly accurate predictions that can ease the workflow in a clinical setting.