Analysis of Granular Parakeratosis Lesion Segmentation: BCE U-Net vs SOTA

Abstract

AbstractAutomatic extraction of the skin lesion is critical in assisting dermatologists in the detection of dangerous skin lesions. Some skin lesions are so small that doctors may not notice them with the naked eye, and dermatologists may struggle to classify them. Nowadays, bio-medical study of lesion images using computational tools is quite widespread. Automated lesion segmentation plays a vital role in the diagnosis of skin diseases like granular parakeratosis. It appears as red or brown papules on the skin and should be treated early. In this paper, the proposed U-Net with binary cross-entropy method is described and compared with several existing segmentation methods, including fully convolutional network (FCN), SegNet and DeepLabv3 + on the lesions of granular parakeratosis. It also presents a comparative study of binary cross-entropy (BCE)-based U-Net with several other state-of-the-art (SoTA) techniques. The images are trained end-to-end, and the performance is evaluated using accuracy, sensitivity and specificity metrics. When compared to existing segmentation algorithms, the U-Net with binary cross-entropy has the greatest performance in terms of all three assessment metrics, accuracy, sensitivity and specificity.KeywordsAccuracyBinary Cross-entropySegmentationSensitivityU-Net