Skin Lesion Classification via Dual-Weighted Learning
Imbalanced medical image classification using MobileViT v2 with dual-weighted learning — inverse square-root sampling + ENS loss — and Grad-CAM heatmaps for clinical interpretability.
Problem
Skin lesion datasets are severely class-imbalanced, causing standard classifiers to be biased toward majority classes while missing rare but high-risk conditions.
Approach
Dual-weighted learning: inverse square-root class sampling for batch composition, plus ENS (Effective Number of Samples) loss to balance gradient contributions. Grad-CAM for visual explanations.
Architecture
MobileViT v2 backbone → classification head. Trained with dual-weighted loss. Grad-CAM attention maps overlaid on input images for interpretability.
Results
0.92 weighted F1 — strong balanced performance across all lesion classes.
Lessons learned
ENS loss and sampling reweighting are complementary: sampling balances what the model sees, ENS loss balances how it learns. Together they substantially outperform either alone.
Next
Fed-BLEND — Federated Conformal Prediction for VLMs
Novel federated conformal prediction method that mitigates hallucinations in federated fine-tuned vision-language models via abstention.
GCTAF — Time-Series Forecasting & Flare Risk Classification
Attention-based forecasting of magnetic field trajectories combined with supervised contrastive learning for solar flare classification.