Back to selected work

Medical-image representation learning

ChestX-ray14 Distillation

Multi-label thoracic disease classification with teacher-student knowledge distillation.

PyTorchtimmRAD-DINOBioViL-TDenseNet-121Vision Transformers

Problem

Chest radiograph classification requires robust multi-label models and a defensible comparison across architectures.

Constraints

Evaluation must separate model architecture, split methodology, and teacher/student training behavior.

Approach

Implemented and evaluated DenseNet-121, ConvNeXt-Tiny, BioViL-T, and Vision Transformer variants, including XrVit and RAD-DINO, with knowledge-distillation pipelines on NIH ChestX-ray14.

My contribution

My work covered architecture implementation and evaluation, training pipelines, and comparison of distilled and direct-supervision approaches.

Results

The exact repository and reproducible metrics are being reconciled before publication on this portfolio.

What I would do next

Add verified split metrics, code/notebook links, and a concise error analysis.