qZACH-ViT enables efficient, interpretable medical AI with INT8 quantization

Compact medical-image classifiers must balance efficiency with interpretability, but these goals are often treated separately. In a new arXiv paper, Athanasios Angelakis introduces qZACH-ViT, a quantization-aware extension of the zero-token, position-free ZACH-ViT backbone. The model incorporates Recursive Attribution-Stabilized Optimization (RASO), which aligns classification and attribution gradients while removing conflicting components. Evaluated across 280 runs on seven MedMNIST datasets with only 50 training images per class, all 210 qZACH-ViT checkpoints were converted to executable mixed-precision INT8 ONNX graphs featuring 16 signed INT8 MatMulInteger projections with INT32 accumulation.

Results show that mixed-precision INT8 qZACH-ViT with Adam improves the FP32 baseline mean on all seven datasets (mean paired gain of 0.0313), while qZACH-ViT with RASO yields a mean gain of 0.0368. Across 964,920 source-to-INT8 test comparisons, prediction agreement is 99.9751%, with a mean absolute primary-metric change of just 0.000133. Attribution maps maintain high fidelity: mean cosine similarity of 0.999955, rank correlation of 0.9944, and top-10% overlap of 0.9692. The ONNX artifacts are 70% smaller than source checkpoints and provide 1.41x and 2.39x end-to-end CPU speedups with one and four threads. RASO significantly reduces sufficiency error and improves input-noise stability over standard Adam with the same attribution loss, establishing qZACH-ViT as a deployable compact intrinsically explainable model.