Much of the molecular activity within tissue remains invisible to the human eye, hidden beneath the morphology captured in routine histology slides. Ongoing projects explore how deep learning can reveal that hidden layer, linking visual tissue structure to the gene expression and cell identities that underlie it. One line of work predicts spatial gene expression directly from H&E-stained images, learning morphological patterns that correlate with molecular activity, while emphasizing lightweight, computationally efficient architectures made practical through model scaling and knowledge distillation. A second project addresses a related gap: since spatial transcriptomics can only measure a limited number of genes, it integrates single-cell RNA-seq data with spatial data to infer the many genes left unmeasured. Rather than relying on the linear mappings or one-to-one cell alignments used in prior methods, this work uses deep learning to capture the complex, nonlinear relationships between modalities while preserving the biological structure of distinct cell populations. A third project extends this idea to cell identity itself, combining gene expression and histological morphology to classify cell types and trace how they relate to disease progression and its molecular markers. It further tests whether embeddings learned from breast cancer cell types can generalize to identify immune cell types in ovarian cancer, probing the transferability of these models across tissues and organs.
Dual-color Dual-hapten In Situ Hybridization (D-DISH) is a computational pathology project focused on making breast cancer diagnosis faster and more reliable by automating one of its most tedious steps, reading HER2 biomarker signals from tissue slides. Right now, pathologists have to manually count tiny signals under a microscope to determine a patient's HER2 status, a process that's slow, exhausting, and can vary from one pathologist to another. This project builds an AI system that looks at these slides, finds individual cells, and detects and counts the relevant signals inside them, automatically calculating the ratio doctors use to guide treatment decisions with the long-term goal of making this tool usable directly in clinical workflows.
Active learning for medical image segmentation must decide which unlabeled samples are most valuable to annotate under a limited labeling budget. Existing strategies such as uncertainty, diversity, and gradient-based acquisition often rely on a single acquisition principle or fixed combinations, which may not adapt well as the model and unlabeled pool evolve across active-learning rounds. We address this by dynamically combining multiple complementary acquisition strategies through a cyclical, adaptive mixture, allowing the contribution of each strategy to change according to its observed performance during active learning. We aim to balance exploration of diverse samples and exploitation of informative uncertain samples, with the goal of achieving stronger segmentation performance with fewer annotations
Intraoral Periapical (IOPA) radiographs are the clinical standard for assessing caries depth and guiding decisions on pulp therapy versus restorative treatment. A key parameter in this decision is the Minimum Remaining Dentin Thickness (MRDT), the distance between the deepest extent of caries and the pulp chamber. Currently, MRDT assessment relies on subjective visual estimation by clinicians, leading to inter-observer variability and inconsistent treatment planning, particularly in resource-constrained settings. This project develops a deep learning pipeline to automate MRDT estimation directly from IOPA radiographs, aiming to provide objective, reproducible measurements that support clinical decision-making.
Fluorescence in situ hybridization (FISH) is a cytogenetic technique which is used to visualise and locate specific DNA sequences in a cell nucleus using fluorescent probes. This enables detection of chromosomal abnormalities which are critical for cancer diagnosis, prognosis and targeted therapies. Usually in clinical labs, this process is carried out by trained clinicians where they evaluate hundreds of nuclei per case under a fluorescence microscope, each taking few minutes depending on the genomic complexity, for assessing signal pattern for mutations, deletions, amplifications, translocations, etc. This manual labour is time consuming, labour intensive, prone to errors and suffers from inter-observer variability, also making it unscalable for large clinical labs. Our work focuses on developing a framework for fully automating this manual signal counting to assist clinicians and researchers in the analysis of complex genomic imaging data.
The Investigative Centre of Advanced Research in Endotypes of Type 1 Diabetes (i-CARE-T1D) is a major national initiative spanning seven centers across India, building India's first structured framework for classifying Type 1 Diabetes into distinct biological endotypes. Type 1 Diabetes doesn't look the same in every patient. In Indians, the disease is often more severe and follows a different course than in Western cohorts, on whose data most existing models are still built. i-CARE-T1D is working to change that, bringing together clinical, genetic, immunological, nutritional, and gut-microbial data from newly diagnosed patients and their siblings across all partner centers. Our work within this initiative turns that data into AI models that characterize these endotypes and predict disease progression in newly diagnosed and at-risk patients, working toward more personalized, precision-medicine-based care for Type 1 Diabetes in India.
Stenosis detection, if achieved in time, can prevent major cardiac disorders like heart attack. Modern day stenosis detection demands significant amount of clinician's time for analysing a single patient's angiogram. The growing time burden, especially against rising patient volumes, highlights the need for a reliable AI assistant that can support clinicians with faster, more consistent diagnosis. The ultimate goal is to shorten the time between imaging and treatment decision, which is often the deciding factor in preventing major cardiac events.
A common observation across major segmentation methods is a systematic bias toward False Negatives (FN) over False Positives (FP), stemming from both data-driven factors and method formulation choices. Rather than addressing FPs and FNs as isolated problems, we hypothesize and empirically demonstrate that explicitly balancing the two during training yields more robust and accurate segmentation outputs. This reframing shifts the objective from minimizing error types independently to optimizing their trade-off jointly, leading to improved boundary delineation and more reliable downstream performance.
Complete Blood Count (CBC) is among the most basic clinical tests. This test is the first indication of any haematological disease and invokes further complex tests. The aim of this project is to aid the clinicians in choosing the right conclusive test by predicting the expected haematological disease using an AI model. While getting adequate amounts of CBC data for this task is relatively simpler, acquiring labels for each sample is difficult due to requirements of multiple subsequent tests for correct disease identification. The challenge lies in using majorly unlabelled data with a small set of labelled samples for disease identification. The catch is that the classes provided in labelled set are not exhaustive and may expand or change overtime.
Deepfake generation and identification have profound applications in the field. However, most of deepfake generation and detection is centred around medical images only. We aim to explore coherent image and report deepfakes from both generation and detection perspectives.