Omid-Nejati/MedViTV2

Integrates Kolmogorov–Arnold Network (KAN) layers within transformer blocks and proposes Dilated Neighborhood Attention (DiNA) to expand receptive fields and prevent feature collapse in corrupted medical images. Supports evaluation across 17 medical datasets plus corruption-robust variants, with pre-trained weights available across four model scales (tiny/small/base/large). Compatible with PyTorch 2.5 and TIMM models, enabling flexible training and inference with Grad-CAM visualization support.