A comparative study of deep learning-based segmentation and classification techniques for melanoma detection

Abstract

Melanoma, a perilous disease, exerts significant pressure on global healthcare systems, and although early identification can reduce mortality rates, conventional diagnostic methods often fail to accurately detect this disease due to its intricate nature. Having a convenient means to reach an automated and sturdy system that can detect melanoma by analysing a dermatoscopic image of lesions can prove to be an invaluable asset in the realm of medicine. Deep Learning based on Convolutional Neural Networks is a modern approach used in computer-assisted medical diagnosis, specifically for developing systems that focus on segmentation, classification, and detection of melanoma. Segmentation techniques, which have the potential to improve performance, are underexplored in the state-of-the-art methods for automated medical diagnosis. This paper aims to assess the efficacy of deep learning -based convolutional neural networks techniques for the purpose of automatically segmenting skin lesions based on dermatoscopic images. The approach outlined in this research paper employs a two-step process: the initial step employs the U-Net design and the Watershed-based Convolutional Neural Network method to automatically isolate the area of interest within a dermatoscopic image, while the second step involves categorizing lesions using deep learning architectures such as ResNet50 and EfficientNetB3. The effectiveness of the approach was extensively assessed by utilizing established performance metrics such as accuracy, specificity, sensitivity, and Area Under the Curve (AUC). Among the proposed models, U-Net with ResNet50 model has performed well compared to Watershed and EfficientNetB2 models with accuracy of 88%. This model also has outperformed other state-of-the art ResNet models.