SKIN LESION CLASSIFICATION USING DEEP LEARNING

Abstract

Skina alesion aclassification is a critical task in dermatology afor athe early adetection and diagnosis of askin acancer. With the advancements in adeep alearning, models such as EfficientNet and Vision Transformer have demonstrated remarkable performance in various computer vision applications. This thesis investigates the effectiveness of EfficientNet and Vision Transformer models for askin aalesion aclassification. The aprimary aobjective is to assess the performance of these models in accurately identifying different types of skin lesions, including melanoma, keratinocyte cancers, and other benign or malignant conditions. The research methodology involves training the amodels on a acomprehensivea adataseta of annotated skina lesiona aimages obtained from diverse sources. The images are preprocessed and augmented to enhance the robustness and generalization of the models. Extensive experimentsaa are aconducteda to aevaluate and compare the performance of aEfficientaNet and Vision Transformer models with traditional computer vision approaches, considering metrics such as aaccuracya, aprecisiona, arecalla, and aaFa1 score. Furthermore, the efficiency and scalability of the models are examined, considering their potential deployment in real world clinical settings. The research also investigates the ainterpretabilitya of athe amodels to gain insightsa into the featuresaa and apatterns alearned for askin alesiona classificationa. The findings of this research contribute to the advancement of computer-aided diagnosis systems in dermatology. The results showcase the potential of EfficientNet and Vision Transformer models in improving the accuracya and aefficiency of askin alesion classificationa. This research provides valuable ainsights and aguidelines for adermatologists and aresearchers aworking in the afield of askin acancera adetection, ultimately aiding in the early diagnosis and treatment of skin cancer for improved patient outcomes.