A Unified Consistency-Calibrated Boundary-Aware Framework for Generalizable Skin Cancer Detection

Abstract

Skin cancer detection by automated methods faces significant challenges in generalizing across diverse patient populations. This limitation is due to the wide variability in the appearance of skin lesions and the lack of standardization in dermatological imaging. To address this issue, this research paper proposes a stability-scale calibrated threshold framework (C2BASC++), which is designed to improve robustness and diagnostic accuracy in hospital settings. By integrating threshold-sensitive feature extraction and a stability measurement process, this framework significantly improves the accuracy and robustness of detection. Experimental results confirm its capabilities in defining disease thresholds, reducing false positives, and robust cross-dataset generalization. The proposed C2BASC++ framework sets a new benchmark in skin disease segmentation, and achieves 96.2% peak IoU and 98.5% specificity in benchmarks against models such as HPO-MMSS and Vision Transformer. Its consistent scalability is a key finding; regardless of the dataset size, C2BASC++ maintains a 6-7% performance gain over baseline models, making it a more reliable and data-efficient solution for automating skin cancer detection.