I built the medical-image-classification-transformers repository to evaluate whether MAE‑generated synthetic images can address class imbalance and limited labeled data in medical imaging. The full workflow – from self‑supervised pre‑training with a Masked Autoencoder to downstream classification with a DeiT transformer – lives in a single reproducible Jupyter Notebook.
The project started 2026-01, last updated 2026-01.
My goal is to address class imbalance and limited labeled data in medical imaging.
I use the DermaMNIST multi‑class skin‑lesion dataset, which has an inherently imbalanced class distribution.
I employ self‑supervised learning using Masked Autoencoders (MAE).
MAE is used to reconstruct and generate synthetic samples.
DeiT (Data‑efficient Image Transformer) serves as the downstream classifier.
Baseline (No Augmentation) Accuracy = 0.757.
MAE‑Augmented Accuracy = 0.762.
Baseline Macro‑F1 = 0.54.
MAE‑Augmented Macro‑F1 = 0.57.
Baseline Weighted‑F1 = 0.76.
MAE‑Augmented Weighted‑F1 = 0.77.