DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis Colloquium
Speaker : Mr. Archakam Satvikam Anudeep
S.R. Number : 06-18-01-10-22-23-1-23183
Title : “Improving Out-of-Distribution Robustness in Long-Tailed Medical Image Datasets”
Research Supervisor : Dr. Vaanathi Sundaresan
Date & Time : December 02, 2025, 10.00 AM
Venue : # 102 CDS Seminar Hall
ABSTRACT
Deep neural networks (DNNs) have achieved remarkable success across medical imaging applications. However, their clinical deployment remains limited due to vulnerability to out-of-distribution (OOD) data, i.e., inputs that differ from the training distribution, leading to overconfident yet incorrect predictions. This challenge is further compounded by class imbalance and domain shifts in real world clinical data. While long tailed OOD detection has been explored, prior work mostly addresses natural images or narrowly targets medical cases such as skin lesions, with restricted evaluation on novel classes.
To address these gaps, we propose a unified long-tailed OOD detection framework for medical imaging, designed and evaluated under clinically realistic conditions. At its core lies a novel von Mises Fisher (vMF) based classifier that learns nonlinear decision boundaries, enabling more flexible and robust feature separation than conventional vMF formulations. To mitigate bias toward head classes in long-tailed distributions, we employ a multi-expert design where several margin-adjusted vMF classifiers specialize across different regions of the label space. Their aggregated predictions reduce imbalance and enhance generalization. In addition, an outlier expert trained with an outlier exposure objective captures atypical feature patterns and suppresses overconfidence on unseen data. Together, these components form an ensemble that balances class representation and strengthens OOD robustness.
Furthermore, our work advances long-tailed OOD detection in three key aspects. We train and evaluate our framework on three diverse medical imaging datasets—ISIC2019 (dermatology), NIHCXR (chest radiography), and RFMiD (retinal imaging)—whereas most prior studies focus solely on skin lesions. We further introduce a clinically reflective OOD evaluation spectrum that spans near OODs (e.g., corrupted images and unseen disease variants) to far OODs (cross-modality and non-medical domains), providing one of the first systematic assessments of distributional shifts encountered in clinical deployment. Within this realistic setup, we benchmark multiple state-of-the-art methods and demonstrate that our proposed approach achieves the best overall performance. Compared to state-of-the-art methods, our method reduces the average FPR95 by 10.95%, 18.73%, and 7.68% on ISIC2019, NIHCXR, and RFMiD respectively, while outperforming the baseline by over 59% across all datasets. In addition to enhanced OOD detection, our framework attains the highest balanced accuracy for ID classification, underscoring its effectiveness in learning robust decision boundaries under long-tailed conditions.
ALL ARE WELCOME
