MAE-Model-MedMNIST-Predictor

Using Masked Autoencoder model pretraining to Predict Medical Imaging Results.

For Code of the Model Architecture and Training:

Github Model Code

For Live Demo:

Hugging Face Model Space Demo

Results:

The best predictor variant was the cross attention pooler with 95.11% test accuracy.

This test accuracy beats the best ResNet (ResNet-50 (28)) model accuracy from the official MedMNIST v2 dataset benchmark by 4%.

Here are the variants and their test accuracies and AUCs:

Variant	Accuracy	AUC
Benchmark ResNet-50 (28)	0.911	0.990
Linear Probe (GAP)	0.9415	0.9942
Cross Attention Pooler	0.9511	0.9942
Cross Attention Hybrid	0.9433	0.9954