Automated Skin Cancer Screening Using High-Performance Convolutional Neural Networks and Discrete Cosine Transform

Skin cancer is a potentially fatal disease that requires early and accurate detection to enable successful treatment. Visual inspection techniques frequently require accuracy enhancement to improve clinical decision-making and increase the probability of successful treatment outcomes. This paper presents the development and evaluation of high-performance deep learning (DL) models combined with the discrete cosine transform (DCT) for automated skin cancer classification and screening for early skin cancer diagnosis. Both methods are used to extract features from medical images. The proposed method starts with DCT feature reduction and then uses convolutional neural networks (CNNs) DL models for classification. Pretrained DCT-VGG-16 and DCT-ResNet50 models are employed for high-accuracy classification with low computational cost. The performance is rigorously evaluated on a large dataset of ISIC 2019 dermoscopy images. It uses transfer learning with the two pre-trained architectures with metrics including training loss, training accuracy, validation loss, validation accuracy, sensitivity, and specificity for identifying malignant lesions. The ISIC 2019 dataset, comprising 25,331 dermoscopy images across eight diagnostic categories, facilitated robust model training and validation. Findings show that DCT - VGG 16 outperforms DCT-ResNet-50 in accuracy and sensitivity; however, the training time of DCT -ResNet-50 is lower than that of DCT-VGG16. DCT-VGG16 achieved a higher accuracy of 99.4% and sensitivity of 97. 6%; however, DCT -ResNet-50 achieved a lower accuracy of 96.5% and sensitivity of 95.1 %. Based on the outcomes, dermatologists could find the results helpful for early automated skin cancer diagnosis and better patient care.