Digital Pathology & Clinical AI
Cardio-Oncology Echocardiography Cine-Loop Analysis (EchoRisk · MICCAI 2026)
Single-encoder approach for LVEF estimation, left ventricular dysfunction classification, and early cardiotoxicity prediction from echocardiography cine loops. Placed 2nd (early cardiotoxicity prediction), 3rd (LV dysfunction classification), and 8th (LVEF estimation) in the EchoRisk Challenge at MICCAI 2026.
Glioma Histologic Sub-region Classification (BraTS-Path)
Rare-class conditional routing with foundation model representations for glioma histologic sub-region classification. Published at MICCAI 2026 Workshops and Challenges.
Native 3D Self-Supervised DINOv2 for Brain MRI
Scaling native 3D self-supervised DINOv2 pretraining to build brain MRI foundation models. Published at MICCAI 2026 Workshops and Challenges.
Prospective Evaluation of AI-based Glioma Classification
Reproducible HPC-based tool for WHO 2021 adult-type diffuse glioma classification from H&E slides, now in prospective validation in hospital workflows. Presented at SNO 2026.
Vision Language Model for Pathology Report Generation (REG 2025)
Developed a vision-language model to automatically generate structured diagnostic reports from gigapixel whole-slide images. Submitted to the MICCAI 2025 REG challenge, achieving 5th place internationally; extended with autoregressive and retrieval-based chain-of-thought reasoning (MICCAI 2026 Workshops).
WHO 2021 Glioma Classification from WSI
Multi-scale whole-slide image assessment pipeline for automated WHO 2021 glioma subtype classification using pathology foundation models and Multiple Instance Learning (MIL). Trained and validated on 2,500+ brain tumor cases from >25 clinical sites. Published in Neuro-Oncology. Won Best Paper at CCBB 2025.
IDH Mutation Biomarker Identification in Gliomas
Weakly supervised computational algorithm for IDH mutation status prediction from H&E-stained WSI using Self-Supervised Learning (SSL) based on Vision Transformers (ViT) and DINO. Eliminates dependency on expensive molecular testing. Published in Neuro-Oncology Advances.
Prognostic Risk Stratification of Gliomas
AI-based prognostic risk stratification in adult-type diffuse gliomas leveraging pathology foundation models and MIL architectures. Combines WSI morphology with survival prediction objectives for clinically actionable outputs.
Multimodal Prognostic Stratification of Glioblastoma (UPenn · CBICA)
Multimodal explainable AI model integrating histopathology imaging and clinical features for GBM patient prognostic stratification. Achieved AUC ≈ 0.75 on a multi-institutional test cohort of 188 GBM patients. Published in Modern Pathology and Frontiers in Neuroscience.
Oncocytic Renal Tumor Subtyping
Computational algorithm for AI-based subtyping of oncocytic renal tumors (Renal Oncocytoma, Low-Grade Oncocytic Tumor, Chromophobe RCC) from H&E-stained WSI. Published in Archives of Pathology & Laboratory Medicine.
Computer Vision & Prior Work
Skin Lesion Segmentation & Classification (ISIC Challenge)
GAN-based architecture for skin lesion segmentation combined with CNN-based classification using ResNets, MobileNets, Xception, and EfficientNets. Published in Nature Scientific Reports.
Retinal Fundus Glaucoma Analysis (REFUGE2 · MICCAI 2020)
Algorithm for optic disc/cup segmentation, glaucoma classification, and fovea localization from retinal fundus images using EfficientNets and UNet architectures. Ranked Top 10 internationally at MICCAI 2020.
Anomaly Pattern Segmentation in Agriculture (Agriculture-Vision · CVPRW 2020)
Feature Pyramid Network (FPN) encoder with varying decoder backbones for aerial agricultural image anomaly pattern recognition. Results published at CVPR 2021 Workshops.
Semantic Scene Segmentation · Indian Driving Dataset (AUTONUE)
Encoder-Decoder architecture (InceptionResNetV2, EfficientNets) for semantic scene segmentation in unstructured driving environments. Won 1st Prize at NCVPRIPG 2019 and 2nd Prize at ICVGIP 2020.
Automatic Number Plate Recognition (Nymo.ai)
End-to-end real-time ANPR system combining YOLO object detection and Mask-RCNN Region Proposal Networks with OCR integration. Delivered real-time inference at ~6-8 fps on edge devices.
Age-Related Macular Degeneration Detection (AMD Challenge · ISBI 2020)
CNN-based algorithms for AMD diagnosis and lesion segmentation from fundus photos. Ranked Top 10 internationally at IEEE ISBI 2020.