Abstract
This paper introduces a continual learning framework for vision-language models in toxicologic pathology that addresses AI safety and alignment challenges in pharmaceutical drug development, where misalignment between model predictions and biological mechanisms can have severe consequences. While vision-language models have demonstrated remarkable zero-shot transfer capabilities in general computer vision tasks, their application to specialized medical domains like toxicologic pathology remains limited by the static nature of conventional fine-tuning approaches. By developing novel extensions of Low-Rank Adaptation to decompose weight updates into principal and residual subspaces, our method preserves general pathology knowledge, which is critical for safe, foundational reasoning, in the principal subspace while confining toxicologic-specific adaptations to the residual subspace. A gradient-projection-based LoRA initialization ensures new task gradients are orthogonally aligned with prior knowledge, preventing catastrophic forgetting and maintaining trustworthiness as models evolve alongside scientific advancements. This represents the first continual learning approach for VLMs in toxicologic pathology, offering a scalable, parameter-efficient solution for regulatory and drug discovery applications where AI systems must adapt safely to new toxicologic findings without compromising prior safety-critical insights during non-clinical safety evaluation stages.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_056.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=%7EArijit_Patra1
BibTex
@InProceedings{PatAri_Continual_MICCAISAT2026,
author = { Patra, Arijit AND Dandapath, Iman AND A.S., Anudeep AND Sundaresan, Vaanathi},
title = { { Continual Learning with subspace adaptation in foundation models for drug safety studies } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}