FDUM-Net: An enhanced FPN and U-Net architecture for skin lesion segmentation

Abstract

Early detection is essential for the successful removal of all malignant lesions from the body, and skin cancer is one of the most widespread cancers globally. In medical image analysis, identifying the diseased area or the region of interest (ROI) significantly relies on advanced network models. Segmenting skin lesions is a strenuous task due to the presence of varied lesion shapes, ambiguous edge borders, low contrast, and presences of artifacts and noises. Performing manual identification of ROI on a large-scale skin lesion assessment is challenging. This study proposes enhanced FPN and U-Net network models for supervised skin lesion segmentation. The study investigates eight Convolutional Neural Network architectures, including U-Net (classic), U-Net + MobileNet, U-Net + InceptionV3, U-Net + DenseNet121, FPN(classic), FPN + MobileNet, FPN + InceptionV3, and FPN + DenseNet121. The performance of these architectures is evaluated using three optimizers (RMSProp, Adam, and SGD) on the ISIC 2016 dataset. The evaluation metrics include accuracy, IoU, and Dice coefficients on the testing dataset. The experimental findings demonstrate that the FPN architecture with DenseNet121 as the backbone encoder and the U-Net architecture with MobileNet as the backbone encoder achieved the highest dice coefficient of 0.93, accuracy of 0.96, and IoU of 0.87. Our proposed solution for enhancing skin lesion segmentation is called FDUM-Net, which is a combination of enhanced FPN with DenseNet as encoder and U-Net with MobileNet designed to capture high-level information and context for more accurate results. These outcomes surpass the performance of previous research and can assist dermatologists in diagnosing skin cancer more efficiently.