Robust Patient-Level Evaluation of Post-Training Quantization for Edge-Deployable MPox Skin Lesion Classification

Abstract

In this study, we implemented Post-Training Quantization (PTQ) for Mpox skin lesion classification using a patient-level evaluation protocol based on a stringent pre-registered experimental design. The publicly available MSLD v2.0 dataset was preprocessed at the pixel level, resulting in 755 unique original images from six diagnostic classes representing 524 patients. A stratified patient-level split (80% training, 10% validation, and 10% testing) was applied to prevent data leakage. Each experiment was repeated across three independent data splits and three training seeds to assess performance variability and robustness. A two-stage transfer learning framework based on MobileNetV2 was trained and evaluated using FP32, Float16 PTQ, and full INT8 quantization implemented through TensorFlow Lite (TFLite). The FP32 baseline achieved an accuracy of 69%, a macro F1-score of 69%, and a macro-AUC of 94.7%, with a model size of 4.27 MB. Float16 quantization maintained comparable predictive performance and model size, whereas full INT8 quantization reduced the model size to 2.59 MB, representing a 39% reduction compared with the FP32 baseline. Top-2 accuracy remained stable at approximately 89% across all precision regimes, supporting the feasibility of deployment in triage-oriented clinical settings. Variability analyses revealed that performance fluctuations were driven primarily by patient partitioning rather than reduced numerical precision, demonstrating that INT8 quantization preserves clinically relevant performance while substantially reducing computational and storage requirements for deployment in resource-constrained edge environments.