EARLY DETECTION AND MULTI-CLASS CLASSIFICATION OF SKIN CANCER USING DEEP LEARNING AND TRANSFORMER-BASED ARCHITECTURES

Abstract

Skin cancer is one of the most common cancers across the world and there is no sign of a decline in its rate of incidence which is only rising due to various factors including ultraviolet exposure and lifestyle changes. Early detection is an important step in improving survival rates of patients with melanoma as well as complexity of the treatments; however conventional methods of diagnosis such as visual inspection and dermoscopy depend heavily on expert interpretation and are susceptible to interobserver variability. In the past years, deep learning-based methods have become powerful means for automated medical image analysis, which can directly learn complex and discriminative features from the raw image data. This work focuses on a general deep learning system for the early detection and the multiclass classification of skin cancer from dermoscopic images. The experimental analysis consists of traditional support vector machine (SVM) with handcrafted features, convolutional neural network (CNN) architectures ResNet18, DualBranch ResNet18, DenseNet121, EfficientNetB0 and an advanced transformer-based model using a global contextual information within skin lesion images. Model performance is evaluated on a tight evaluation protocol in terms of accuracy, macroaveraged values of precision, recall and F1 score, balanced accuracy, agreement measures, and receiver operating characteristic analysis. The comparative results show a marked improvement with teacher learning from the traditional machine learning techniques to deep CNNs such that the transformer-based architecture gives the most consistent and robust classification results when considering all diagnosis categories. The findings illustrate the success advanced deep learning models can achieve in reducing class confusion especially for visually similar lesion types and raise possibilities of the approach to aid in clinical decision making. Overall, this work gives empirical proof that modern deep learning and transformer based approaches can considerably improve the reliability and accuracy of computer aided detection systems of skin cancer, hence opening a promising direction for future computer aided diagnostic tools in dermatology.