Abstract
Anywhere on the body might develop melanoma, a very serious form of skin cancer. Early detection of melanoma lesions significantly increases the likelihood of effective therapy. In recent times, learning-based segmentation techniques have outperformed conventional algorithms in the segmentation of images. This work presents a novel approach to improve the identification and classification of skin lesions with cancer. We propose a two-stage procedure based on deep learning models. To evaluate our approaches, we used the popular ISIC 2018 dataset, which is well-known for its Skin Lesion Analysis Towards Melanoma Detection Challenge. There are two primary parts to the suggested methods for segmenting and identifying lesions in real time. First, we use an enhanced version of You Only Look Once edition 8 (FYOLOv8) to accurately localise skin lesions. We next use the updated Segmentation Network (F-SegNet) to handle the F-YOLOv8 data further. We conducted experiments on 20,250 photos from three publicly available datasets: The International Skin Imaging Collaboration (ISIC) 2019, International Symposium on Biomedical Imaging (ISBI) 2017, and PH2. The findings were promising. The suggested approach obtained accuracy of 98.50% and 98.50% on the PH2 and ISBI 2017 datasets, respectively, and a Jac score of 93.22% on the ISIC 2019 dataset. In most situations, our technique showed somewhat better performance compared to current efforts in this field utilising predetermined parameters.
Keywords: Melanoma, Skin Cancer, Skin Lesion Segmentation, YOLO.