An Integrated Uncertainty Based Deep Learning Framework for Melanoma Diagnosis and Lesion Segmentation

A dangerous kind of skin cancer that may develop anywhere on the body is melanoma. Early identification of melanoma lesions greatly improves the likelihood of successful therapy. In picture segmentation, learning-based segmentation techniques have recently surpassed conventional algorithms. A deep learning-based classification and Triangular Intuitionistic fuzzy-based segmentation framework is introduced in this paper to enhance the identification and categorization of malignant skin lesions. The Skin Lesion Analysis towards Melanoma Detection Challenge dataset is used to test the framework. In order to segment and detect lesions in real time, the approach consists of two primary processes. Preprocessing involves K-Medoid Clustering-based noise detection followed by Fuzzy Logic-oriented noise removal to enhance image quality and remove unwanted artifacts. After preprocessing, skin lesions are precisely localized using You Only Look Once edition 8 (YOLOv8). There are two steps in the segmentation process to determine the affected areas. Phase I computes the smallest spanning tree for identifying impacted zones based on threshold values using a graph-based framework. In the phase II, two segmentation approaches—Triangular Intuitionistic Fuzzy Numbers (TIFNs) and Triangular Dense Neutrosophic Numbers (TDNNs)—are comparatively evaluated. The comparative analysis demonstrates that TDNNs offer superior performance, yielding more accurate and reliable segmentation results. The 20, 250 photos used in the experiments are from three public datasets: PH2, ISBI 2017, and ISIC 2019. With a Jac score of 96.13% on the ISIC 2019 dataset and 99% accuracy on the ISBI 2017 and 99% PH2 datasets, the findings were encouraging. The suggested strategy performed marginally better than current frameworks with predetermined settings in the majority of circumstances.