Kushal Borkar supervised by Prof. Jawahar C V received his Master of Science in Computer Science and Engineering (CSE). Here’s a summary of his research work on Continual Learning in Interactive Medical Image Segmentation:

Automated segmentation of medical image volumes holds great promise in alleviating the annotation burden on medical professionals, yet it remains a complex challenge due to modality variability, image intricacy, and limited labeled data. While interactive segmentation methods and foundational models have made strides by incorporating user prompts, they often overlook the sequential and contextual dependencies inherent in 3D medical image volumes. This results in segmentation outputs that suffer from inconsistencies, spatial discontinuities, and a lack of anatomical coherence—issues that diminish the clinical reliability of such systems. To address these limitations, we propose a novel interactive segmentation framework that performs test-time training guided by user-provided scribbles, dynamically updating model parameters during inference. Our approach preserves spatial and contextual information across consecutive slices using a student-teacher learning paradigm, which captures both intra-volume continuity and prior knowledge from the training distribution. Evaluations across diverse datasets—including CT, MRI, and microscopy—demonstrate superior accuracy and efficiency. Our method achieves a Dice score of 0.9 within just 3–4 user interactions, significantly reducing annotation time by up to 6.72× compared to manual methods. Moreover, it generalizes well to unseen objects with minimal user input, offering a robust and anatomically coherent solution for interactive medical image segmentation.

December 2025