Skin cancer continues to be one of the most common and deadly diseases worldwide, and early and accurate detection is essential to increase the rates of survival from skin cancer. However, automated skin cancer diagnosis faces challenges due to class imbalance, inter-class similarities, and limited feature representation. This paper presents a hybrid deep learning framework that utilizes preprocessing, augmentation, and normalization to increase the quality of images used for diagnosing skin cancer and to improve the balance of the datasets. Two light weight deep convolutional neural networks (CNNs) such as EfficientNetB1 and NasNetMobile are utilized to extract features from the dermoscopic images using colour, texture, and edge information. Extracted features are combined utilizing a Squeeze-and-Excitation(SE) attention mechanism followed by lightweight DenseNet-based classifier to provide accurate predictions regarding skin cancer. The proposed model is evaluated on five benchmark datasets: ISIC Kaggle, PAD-UFES-20, ISIC 2018, ISIC 2017, and ISIC 2016, achieving classification accuracies ranging from 96% to 98%, respectively. These results confirm the effectiveness and scalability of the proposed approach for reliable real-world clinical skin cancer detection.
Bhargavi et al. (Fri,) studied this question.