Deployment realism
Test whether continual-learning approaches remain credible on heterogeneous, low-prevalence medical data rather than only on simplified benchmarks.
Research
My work asks a deployment-oriented question: when continual-learning systems fail on real medical images, is catastrophic forgetting the primary problem, or are preprocessing, representation, labeling, and task design limiting the model before forgetting can even be meaningfully assessed?
Research objectives
Method diagnosis
Separate forgetting from failures caused by preprocessing, feature quality, label granularity, and insufficient task learnability.
Pipeline improvement
Develop concrete changes—learnable CT windowing and patient-level multiple-instance learning—instead of only retuning existing methods.
Open problems
Identify limitations that should shape future research in continual medical imaging, including evaluation under scarce data and limited compute.
Related research experience
Spinal Disease Detection via Deep Multimodal Fusion
Lead-author BSc thesis work combining X-ray imagery with demographic features to classify eight spine conditions, followed by segmentation.
Deep Learning for Wound Tissue Segmentation
Comparative evaluation of full-image, patch, and superpixel formulations across a novel six-tissue wound dataset.
Read on arXiv →Survival Prediction in Palliative Care
Multimodal modeling of wearable actigraphy and clinical information for survival-outcome prediction.
Open article →Offensive Bengali Meme Detection
Earlier multimodal work combining image and text models through a weighted ensemble architecture.
Open on IEEE Xplore →