Lossless Mechanistic Compression and Surgical Correction of Medical Imaging Models

Medical imaging models such as CheXNet (DenseNet121) are widely deployed for multi-label thoracic pathology classification but suffer from large parameter counts (6.97M) and opaque debugging pipelines. We present a unified framework that (1) losslessly compresses the model by 51.43% (6.97M → 3.38M parameters) via channel-wise sparsity-constrained weight reconstruction on NIH ChestX-ray14, with AUROC change of +0.0004 and per-image latency reduced from 15.17 to 14.73 ms; (2) enables surgical correction through classifier-channel attribution and selective weight zeroing (5-channel correction reduces a target false positive by ∆prob −0.13 with zero true-positive loss and exactly zero AUROC change on the other 13 pathologies); (3) provides a cost-aware Treatment Decision System routing each pathology issue to its
cheapest effective intervention; (4) supports clinical report auto-generation with channel level evidence, Grad-CAM region mapping, and mutual-exclusivity-based exclusion. We further discover that polarized classifier channels are not architectural conflicts but bipolar discrimina tive axes exploiting label mutual exclusivity (Jaccard < 0.1 in 89 of 100 polarized channels). On CheXNet, per-class Youden’s-J threshold calibration alone (no retraining) raises the cohort average F1 by ×1.6 and recall by ×7 relative to the default threshold-0.5 operating point; classifier-head fine-tuning (18K trainable parameters, ≈ 85 s) adds at most marginal additional F1 and regresses several best-performing classes, so we characterize threshold calibration as the operationally meaningful intervention and fine-tuning as a mean-AUROC stabilizer. Compres sion method specifics are proprietary; model weights, inference code, and analysis scripts are released for reproduction.