Comparative Performance of Pre-Trained CNN Architectures on Dermoscopic Pigmented Skin Lesions Classification

Pigmented skin lesions are skin disorders with increased incidence in the last decade. Early detection and accurate diagnosis can reduce mortality. Deep learning, particularly the Convolution Neural Network (CNN), has reached impressive results for many purposes involving image classification. The use of CNN architecture does not require pre-processing at the level of an image as that of traditional methods. The required pre-processing is only an augmentation to increase the data needed for model training. This work compared ten pre-trained CNN architectures: Inception-V3, ResNet-50v2, ResNet- 152v2, InceptionResNet-v2, DenseNet-201, Xception, MobileNet, MobileNet-v2, NASNetLarge and EfficientNet-B7 to determine their performance in classifying pigmented skin lesions. The dataset is ISIC 2019, with eight disease classes that apply data augmentation and balancing. The architectures were evaluated in a multiclass statistical evaluation based on their accuracy, sensitivity, specificity, precision, F-score, and AUC. The findings of this research those four pre-trained CNN architectures, namely Inception-v3, InceptionResNet-v2, NASNetLarge, and DenseNet- 201, showed superior performance with AUC values above 0.90.