{"id":1178034,"date":"2026-07-07T00:15:04","date_gmt":"2026-07-07T07:15:04","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-project&#038;p=1178034"},"modified":"2026-07-07T07:46:34","modified_gmt":"2026-07-07T14:46:34","slug":"care","status":"publish","type":"msr-project","link":"https:\/\/www.microsoft.com\/en-us\/research\/project\/care\/","title":{"rendered":"CARE"},"content":{"rendered":"<section class=\"mb-3 moray-highlight\">\n\t<div class=\"card-img-overlay mx-lg-0\">\n\t\t<div class=\"card-background bg-gray-200 has-background- card-background--full-bleed\">\n\t\t\t\t\t<\/div>\n\t\t<!-- Foreground -->\n\t\t<div class=\"card-foreground d-flex mt-md-n5 my-lg-5 px-g px-lg-0\">\n\t\t\t<!-- Container -->\n\t\t\t<div class=\"container d-flex mt-md-n5 my-lg-5 \">\n\t\t\t\t<!-- Card wrapper -->\n\t\t\t\t<div class=\"w-100 w-lg-col-5\">\n\t\t\t\t\t<!-- Card -->\n\t\t\t\t\t<div class=\"card material-md-card py-5 px-md-5\">\n\t\t\t\t\t\t<div class=\"card-body \">\n\t\t\t\t\t\t\t\n\t\t\t\t\t\t\t\n\n<h1 id=\"care\" class=\"wp-block-heading\">CARE<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">CARE is a research program at Microsoft Research India focused on building foundation and multimodal models for healthcare, with an emphasis on radiology. The program develops vision-language and representation-learning systems that prioritize clinical fidelity, interpretability, and robustness over surface-level fluency. Work is conducted in collaboration with clinical partners, with validation on real-world data including populations underrepresented in existing radiology datasets.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n\n\n\n<p class=\"wp-block-paragraph\">Through project&nbsp;<a href=\"\/Users\/meranjit\/AppData\/Local\/Programs\/Microsoft%20VS%20Code\/7e7950df89\/resources\/app\/out\/vs\/code\/electron-browser\/workbench\/workbench.html\">CARE<\/a>&nbsp;(Clinically Aligned Radiology Expertise), we investigate how AI assistants can be integrated into radiology workflows in a reliable and clinically meaningful manner. CARE develops foundation and multimodal models for radiology across 2D and 3D imaging, with an emphasis on clinical fidelity, calibration, and interpretability \u2014 aligning model behavior with the way radiologists reason and validating it through collaborations with hospital partners.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The program currently comprises two sub-projects:&nbsp;CARE-X, addressing chest X-ray (2D) interpretation, and&nbsp;MedCompose, addressing compositional representation learning for 3D imaging such as CT and MRI.<\/p>\n\n\n\n<h2 id=\"care-x\" class=\"wp-block-heading\">CARE-X<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">CARE-X is a chest X-ray vision-language model designed to produce outputs a clinician can act on, rather than outputs that merely read well. The chest X-ray is the most common medical image worldwide, and its interpretation involves detecting findings, localizing them, performing measurements, and composing a structured report. While current vision-language models perform the report-writing task competently, a well-formed report can still be clinically incorrect. CARE-X targets three specific failure modes of generative radiology models.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>1. Calibration and thresholdable predictions<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A purely generative model emits diagnoses as free text, providing no probability estimate and no mechanism to adjust the sensitivity\u2013specificity trade-off across clinical settings (for example, screening versus confirmatory use). CARE-X attaches lightweight discriminative heads that share the generative model&#8217;s underlying representation and produce a calibrated, thresholdable score from a single forward pass.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This provides two inference modes from one model:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Generative mode<\/strong> \u2014 autoregressive decoding for free-text reports, VQA answers, and text-based bounding boxes.<\/li>\n\n\n\n<li><strong>Auxiliary-head mode<\/strong>&nbsp;\u2014 discriminative probability scores and detection bounding boxes with associated confidence scores.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The generative and discriminative objectives are complementary: adding the discriminative signal sharpened the shared feature representation and improved generation quality. Clinicians can adjust decision thresholds directly from the model&#8217;s outputs.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>2. Reward-aligned optimization<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Supervised training with cross-entropy optimizes token likelihood, which is a weak proxy for clinical fidelity; a report can achieve high likelihood while omitting a finding or introducing a hallucinated pathology. CARE-X therefore applies reinforcement learning to optimize clinical metrics directly.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For report generation, three complementary rewards are combined so that each compensates for the others&#8217; blind spots:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Entity-level factual correctness.<\/strong><\/li>\n\n\n\n<li><strong>Penalization of omissions and hallucinations<\/strong>&nbsp;of clinically significant findings.<\/li>\n\n\n\n<li><strong>Semantic faithfulness<\/strong>&nbsp;beyond surface wording.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">This reward-driven formulation is extended across all three tasks \u2014 report generation, visual question answering, and localization \u2014 with rewards tailored to what is clinically relevant in each.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>3. Tool-augmented measurement<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A class of diagnoses is defined by measurement against an established threshold rather than by pattern recognition; for example, cardiomegaly is determined by a ratio and is fundamentally a geometric computation. Requiring a neural network to estimate such quantities perceptually is inappropriate. Instead, we need to enable the model to issue tool calls: retaining full visual access to the image, it emits a structured tool call to detect the relevant landmarks, compute the ratio, and compare against evidence-based criteria, then synthesizes the diagnosis from the deterministic result.<\/p>\n\n\n\n<h2 id=\"hidden-causality-in-radiology-reports\" class=\"wp-block-heading\">Hidden Causality in Radiology Reports<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Traditional radiology reports document a patient&#8217;s symptoms, imaging findings, and final diagnosis, but they rarely make explicit the causal relationships and reasoning that lead to that diagnosis. This omission limits both the interpretability of the report and its value for clinical education. As part of the&nbsp;<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/research.nii.ac.jp\/ntcir\/workshop\/OnlineProceedings18\/pdf\/ntcir\/03-NTCIR18-HIDDEN-RAD-RanjitM.pdf\">NTCIR-18 Hidden-RAD Challenge<span class=\"sr-only\"> (opens in new tab)<\/span><\/a>&nbsp;(Hidden Causality Inclusion in Radiology Reports), we investigate how AI models can recover this hidden causality \u2014 generating a causality-exploration section that reflects the diagnostic reasoning a radiologist implicitly performs.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The work also situates domain-specialized small models against a broad set of general-domain and reasoning-focused baselines, examining how effectively causal reasoning can be integrated with a radiology report or image. Beyond the challenge itself, the recovered causal explanations point toward richer, more transparent automated radiology reporting workflows, where a generated report conveys not only what was observed but why a particular diagnosis follows.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.microsoft.com\/en-us\/research\/publication\/hidden-causality-inclusion-in-radiology-reports-with-multimodal-small-language-models\/\">Hidden Causality Inclusion in Radiology Reports with Multimodal Small Language Models &#8211; Microsoft Research<\/a><\/p>\n\n\n\n<h2 id=\"medcompose-compositional-representation-learning-for-3d-medical-imaging\" class=\"wp-block-heading\">MedCompose: Compositional Representation Learning for 3D Medical Imaging<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">MedCompose addresses a foundational problem in representation learning for high-dimensional medical imaging. Modern 3D encoders can process large volumes such as CT and MRI scans and learn dense representations, but these representations tend to be entangled: alongside clinically meaningful features, they encode scanner artifacts, acquisition noise, and patient-specific variation. Such models may perform well during training yet fail to generalize across scanners, hospitals, and imaging protocols because they depend on unstable details rather than semantic structure. The project asks how models can be constrained to represent meaningful clinical concepts rather than noise, and how their reasoning can be made interpretable and aligned with clinical practice.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/rexrank.ai\/ReXGroundingCT\/index.html\">MedCompose@ReXGroundingCT Leaderboard &#8211; Segmentation of Findings from Free-Text Reports<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<h3 id=\"clinical-motivation-compositional-reasoning\" class=\"wp-block-heading\">Clinical motivation: compositional reasoning<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Radiological characterization is compositional. A standard description such as &#8220;calcified, well-circumscribed nodule measuring 5 mm in the lateral segment of the right middle lobe&#8221; decomposes into distinct concept axes:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Textura<\/strong>\u00a0\u2014 internal appearance, density, and composition (&#8220;calcified&#8221;).<\/li>\n\n\n\n<li><strong>Forma<\/strong>\u00a0\u2014 shape and the interface between the lesion boundary and surrounding tissue (&#8220;well-circumscribed&#8221;).<\/li>\n\n\n\n<li><strong>Scala<\/strong>\u00a0\u2014 spatial scale and physical extent (&#8220;5 mm&#8221;).<\/li>\n\n\n\n<li><strong>Locus<\/strong>\u00a0\u2014 anatomical location (&#8220;right middle lobe&#8221;).<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">A finding is thus a structured combination of distinct visual concepts. MedCompose-CT operationalizes this observation through a three-stage architecture.<\/p>\n\n\n\n<h3 id=\"architecture-decompose-discretize-resynthesize\" class=\"wp-block-heading\">Architecture: Decompose, Discretize, Resynthesize<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Decompose.<\/strong>&nbsp;After the 3D encoder extracts visual information, the network branches to separate the feature streams into the four concept axes \u2014 Textura, Forma, Scala, and Locus \u2014 rather than retaining a single mixed representation.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Discretize.<\/strong>&nbsp;Each concept stream passes through a vector-quantized bottleneck, or concept codebook, constraining the model to express observations using a finite vocabulary of recurring visual patterns learned across many scans rather than an unconstrained continuous vector per patient. This encourages the learning of stable, reusable prototypes to which new findings are assigned. The discretization serves two purposes: it improves interpretability by structuring the representation around recognizable concepts, and it provides semantic denoising by discouraging the encoding of scanner-specific noise, acquisition artifacts, and irrelevant patient variation in favor of recurring structure that generalizes across datasets.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Resynthesize.<\/strong>&nbsp;Because clinical meaning arises from the interaction of attributes rather than from isolated traits \u2014 the significance of a small calcified nodule depends on its location \u2014 the quantized concept representations are passed to an attention-based resynthesis module that models the interaction of texture, shape, scale, and location within the 3D scene before producing the final segmentation or report.<\/p>\n\n\n\n<h3 id=\"properties\" class=\"wp-block-heading\">Properties<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Interpretability<\/strong>\u00a0\u2014 predictions are traceable to specific concept axes, allowing inspection of whether the model reasoned from shape or density and whether that reasoning aligns with clinical logic.<\/li>\n\n\n\n<li><strong>Robustness<\/strong>\u00a0\u2014 because the discrete bottlenecks prioritize recurring structure, the model is more resilient to scanner variation, acquisition noise, and protocol differences without explicit domain adaptation.<\/li>\n\n\n\n<li><strong>Compositionality<\/strong>\u00a0\u2014 reusable concept parts allow previously unseen findings to be represented as new combinations of familiar building blocks, supporting generalization.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The approach is intended to generalize beyond 3D medical imaging to any domain describable through structured, reusable semantic concepts, as a step toward more interpretable and faithful visual reasoning.<\/p>\n\n\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>CARE is a research program at Microsoft Research India focused on building foundation and multimodal models for healthcare, with an emphasis on radiology. The program develops vision-language and representation-learning systems that prioritize clinical fidelity, interpretability, and robustness over surface-level fluency. Work is conducted in collaboration with clinical partners, with validation on real-world data including populations [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","footnotes":""},"research-area":[13553],"msr-locale":[268875],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1178034","msr-project","type-msr-project","status-publish","hentry","msr-research-area-medical-health-genomics","msr-locale-en_us","msr-archive-status-active"],"msr_project_start":"2024-10-01","related-publications":[1106766,1156393,1178063],"related-downloads":[],"related-videos":[],"related-groups":[],"related-events":[],"related-opportunities":[],"related-posts":[],"related-articles":[],"tab-content":[],"related-researchers":[{"type":"user_nicename","display_name":"Mercy Ranjit","user_id":43716,"people_section":"Related people","alias":"meranjit"},{"type":"user_nicename","display_name":"Tanuja Ganu","user_id":38883,"people_section":"Related people","alias":"taganu"},{"type":"guest","display_name":"Anirban Porya","user_id":1178035,"people_section":"Related people","alias":""},{"type":"guest","display_name":"Niharika Vadlamudi","user_id":1178039,"people_section":"Related people","alias":""},{"type":"guest","display_name":"Nikilesh Chowdary  Ethamukkala","user_id":1178038,"people_section":"Related people","alias":""},{"type":"guest","display_name":"Prasanth VV","user_id":1178040,"people_section":"Related people","alias":""},{"type":"guest","display_name":"Ananya Shukla","user_id":1178041,"people_section":"Related people","alias":""},{"type":"guest","display_name":"Azad Singh","user_id":1178042,"people_section":"Related people","alias":""},{"type":"guest","display_name":"Amog Rao","user_id":1178043,"people_section":"Related people","alias":""}],"msr_research_lab":[199562],"msr_impact_theme":[],"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/1178034","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-project"}],"version-history":[{"count":11,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/1178034\/revisions"}],"predecessor-version":[{"id":1178182,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/1178034\/revisions\/1178182"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1178034"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1178034"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1178034"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1178034"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1178034"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}