IMPROVED MELANOMA DETECTION USING IMAGE SEGMENTATION AND A LIGHTWEIGHT CNN FRAMEWORK

Abstract

The most severe type of skin cancer is melanoma, and early diagnosis is so important for patients' survival. Manual dermoscopic examination is time-consuming, subjective and impractical for large-scale screening, and existing computational methods have significant drawbacks: handcrafted feature extraction methods do not transfer well between datasets, and deep learning methods like ResNet-50 and Inception-ResNet-v2 achieve higher accuracy but require significant computational resources, unsuitable in the clinic. Furthermore, segmentation and classification are generally considered as two independent processes, and class imbalance in dermoscopic data sets is seldom addressed, and may decrease the accuracy of melanoma detection. In this paper, a two-stage approach deep learning-based automated melanoma detection system is proposed for ISIC 2018 Lesion Segemtation and Classifiation dataset with minimum computational complexity. Preprocessing techniques used were the removal of hair artifacts, the application of a CLAHE contrast enhancement, and pixel normalization; the class imbalance was handled by class augmentation and loss weight. A Mobile-UNet architecture was used to segment lesions followed by a two-stage trained EfficientNet-B0 classification network. Its Dice Coefficient of 87.8%, IoU of 80.6%, classification accuracy of 92.5%, F1-score of 91.8% and AUC-ROC of 95.1% outperform six benchmark methods such as SVM, Random Forest, VGG-16, ResNet-50, Inception-V3, and MobileNetV2-UNet, indicating its feasibility for efficient, real-world clinical utilization.