Comparative evaluation of multiple CNN architectures for dermoscopic skin lesion classification using ISIC dataset

Abstract

Skin diseases pose a significant global health problem, and early diagnosis plays a vital role in improving patient outcomes, particularly for conditions such as Basal Cell Carcinoma, Dermatofibroma, Nevus, and Pigmented Benign Keratosis. This paper presents a systematic comparative evaluation of seven state-of-the-art Convolutional Neural Network architectures - InceptionV3, DenseNet-121, Xception, EfficientNetB3, ResNet152V2, MobileNetV2, and InceptionResNetV2 - for automated classification of dermoscopic images from a 9,857-image ISIC dataset. All models underwent identical preprocessing, data augmentation, and two-phase fine-tuning under controlled experimental conditions to eliminate pipeline bias. InceptionResNetV2 achieved superior performance with 93.5% test accuracy and 0.98 macro-AUC, demonstrating hybrid inception-residual architecture's effectiveness for multi-scale dermoscopic feature extraction. DenseNet-121 followed at 91.6% accuracy, while lightweight MobileNetV2 delivered 91.7% accuracy in 2.1 hours training time, suitable for mobile deployment. Basal Cell Carcinoma detection proved reliable across architectures (F 1=0.76-0.83), though class imbalance impacted Dermatofibroma performance. Confusion matrix analysis revealed Nevus-Pigmented Benign Keratosis overlap reflecting clinical similarity. Results establish clear deployment guidelines: InceptionResNetV2 for hospital diagnosis, MobileNetV2 for rural teledermatology, addressing dermatologist shortages in resource-limited settings like Bangladesh.