Lesion Focus-Lite: Knowledge Distillation For CPU-Deployable Skin Cancer Prediction With Multi-Level Validation

Abstract

Deep learning models for dermoscopic skin cancer prediction are usually reported on one test split and seldom built into software that can run outside a research GPU server. This study implements the Lesion Focus-Net framework, a segmentation-guided hybrid CNN–Transformer design with adaptive pre-processing, and validates it as a working system. The implementation has two tiers. Tier A runs the full framework on a GPU: trigger-based colour constancy, hair removal and contrast enhancement, an attention residual U-Net for lesion segmentation, and a dual-stream Swin-T/EfficientNet-B3 classifier with cross-attention, patient metadata, test-time augmentation and temperature scaling. Tier B, Lesion Focus-Lite, is a 4.2-million-parameter MobileNetV3 student distilled from the full framework, which needs no segmentation and runs through ONNX Runtime on a CPU. Both tiers were trained on a lesion-level split of HAM10000 and validated at five levels: an internal hold-out of 1,975 images, the independent ISIC 2018 Task 3 test set of 1,511 images, robustness to five image degradations, malignant triage with calibration, and a deployment profile. The full framework reached a balanced accuracy of 65.8% and a macro AUC of 0.962 internally, and 71.3% and 0.964 on the independent set, with a malignant-versus-benign AUC of 0.939 and an expected calibration error of 0.035. Distillation raised the light model's internal balanced accuracy from 57.7% to 60.4% and its macro-F1 from 54.2% to 59.4% (McNemar p = 0.008). Lesion Focus-Lite ran end to end in 127 ms per image on two CPU cores, about 23 times faster than the full pipeline, with a macro AUC above 0.94 on both test sets. Pre-processing made all models almost insensitive to colour casts, but Gaussian noise, blur and heavy JPEG compression reduced balanced accuracy by 10 to 36 points, and blur hurt the full framework more than the light model. The results show that the framework generalises to unseen patients, that a distilled model offers a practical CPU option, and that image quality at capture is now the main barrier to reliable use.