Pre-Trained CNN Models for Automatic Classification of Dermis Lession
Saima Bibi, Department of Computer Sciences, University of Engineering and Technology, Taxila, Pakistan.
Javed Iqbal, Department of Computer Sciences, University of Engineering and Technology, Taxila, Pakistan.
Corresponding Author:
Saima Bibi (saima09.uettaxlia@gmail.com)
Abstract:
Skin cancer is a major public health issue, and detecting it early is crucial to enhancing patient outcomes. In this study, we investigate the use of pre-trained deep learning methods for identifying and detecting skin diseases. On a limited dataset of skin lesions dermis, three cutting-edge models namely Xception, EfficientB0, and ResNet101 were applied and assessed. By examining the learnt features at various levels of the CNN network, we used visualization tools to get insights into the model's decision-making process. We found key areas in the input images that contributed to the classification outcomes by using Grad-CAM activation functions. Remarkably, the results demonstrated exceptional accuracy with Xception achieving 94.12%, EfficientB0 93.20%, and ResNet101 86.00%. This outcome underscores the considerable potential of our trained models in the diagnosis of skin cancer.
Keywords:
Transfer Learning; Deep Learning; Feature Learning; Skin Disease.