Medical Computer Vision Lab

Research

My research investigates how medical-imaging systems can learn sequential clinical tasks while preserving useful prior knowledge. My preliminary thesis direction develops modality-aware representation primitives for CT and MRI, and evaluates whether explicitly modeling modality structure improves retention, forward transfer, and task performance in class-incremental learning.

Preliminary thesis direction

MSc thesis · in development

Current direction

Class-Incremental Learning with Modality-Aware Representation Primitives

The thesis asks whether representation primitives that explicitly encode imaging-modality structure can improve continual learning across sequential medical-imaging tasks.

Objective 1: Develop task-conditioned CT representation primitives that adapt to clinically meaningful intensity structure using learned task embeddings.
Objective 2: Design an analogous MRI pulse-sequence conditioning module for T1-weighted, T2-weighted, FLAIR, and DWI data. Candidate datasets include BraTS, AOMIC ID1000, and ADNI, which will be assessed for suitability for sequence-aware and modality-aware experiments.
Objective 3: Integrate the proposed modality-aware representation primitives into class-incremental learning pipelines and compare them with standard continual-learning approaches.
Objective 4: Evaluate whether the proposed representations improve knowledge retention, forward transfer, and performance across sequential medical-imaging tasks.

Diagnostic continual-learning study

Accepted · MICAD 2026

Co-lead & corresponding author

What Limits Continual Learning for Abdominal-Trauma CT Detection? A Diagnostic Study of Forgetting vs. Representation Bottlenecks

Rather than assuming catastrophic forgetting is the only failure mode, this work follows a staged diagnostic sequence: a custom CNN, a pretrained ResNet-18, and patient-level attention-based multiple-instance learning. The experiments compare fine-tuning, EWC, replay, EWC+replay, and LwF under class-incremental and domain-incremental settings.

The central contribution is methodological: continual-learning performance is interpreted only after checking whether the representation, label granularity, and patient-level formulation make the underlying tasks learnable.

Code repository →Publication record →

Research direction

Diagnose before optimizing

Separate catastrophic forgetting from failure caused by weak representations, label mismatch, class imbalance, or insufficient task learnability.

Build modality-aware primitives

Develop representations that explicitly account for CT intensity structure and MRI sequence variation before evaluating continual-learning behavior.

Move toward heterogeneous clinical data

Extend adaptive learning beyond CT toward multimodal settings that combine images with demographic, physiological, and structured clinical information.

Earlier research foundation

All publications →

Published · IEEE · 2026

Spinal disease detection

Lead-author multimodal work fusing X-ray imagery with demographic features for spinal-disease classification.

IEEE Xplore →

Preprint · 2025

Wound tissue segmentation

Dataset creation and implementation supporting a comparative benchmark across full-image, patch, and superpixel formulations.

arXiv →

Published · Cancers · 2023 · Q1

Palliative-care survival prediction

Multimodal modeling of wearable actigraphy and clinical variables for survival-outcome prediction.

Article →

PhD opportunities and research collaboration

If your group works on adaptive medical AI, continual learning, or multimodal healthcare AI, I would be glad to discuss potential research fit or collaboration.

Email (institutional email)Email (personal email)