qZACH-ViT Delivers Quantized, Explainable Medical Image AI

This research introduces qZACH-ViT, an extension of the ZACH-ViT backbone designed for compact, intrinsically explainable medical image classification. A key innovation is its quantization-awareness, allowing for efficient deployment. The model incorporates a zero-token, position-free architecture and a novel Recursive Attribution-Stabilized Optimization (RASO) technique, which aligns classification and attribution gradients to enhance stability. Evaluated across seven MedMNIST datasets, qZACH-ViT with RASO demonstrated improved performance over the FP32 ZACH-ViT baseline. After conversion to mixed-precision ONNX INT8 graphs, the models achieved substantial efficiency gains, being 70% smaller and offering 1.4 to 2.4 times faster CPU inference speeds. Crucially, these efficiency improvements came with minimal loss in prediction accuracy (99.9751% agreement) and high fidelity in intrinsic explanation maps, establishing qZACH-ViT as a deployable solution for explainable AI in medical imaging.