Segmentation and Classification of Dermatoscopic Skin Lesion images using U-Net and MobileNet models

Abstract

The 17th most prevalent cancer in the world is cutaneous melanoma. Early detection and adequate treatment are essential for skin cancer success. It might not be possible to tell benign lesions from malignant tumours just by looking at them. The histopathological study of the skin biopsy is the gold standard procedure. Skin biopsy has some drawbacks, including its invasiveness, the pain it causes, and the requirement for many samples for suspected lesions with multiple presentations. Clinical diagnosis can also be aided by noninvasive tools. Several non-invasive imaging techniques are now available to diagnose melanoma because of numerous scientific and technological developments. The most advanced network for pattern identification in medical image analysis is the convolutional neural network (CNN). Thus, utilizing these advanced techniques a Skin Lesion Classifier can be built, which performs segmentation of the Lesion area in the pre-process step to avoid extracting and remembering other additional features from the background of the dermatoscopic image. For the segmentation task U-Net Architecture is utilized on the PH2 dataset and obtained validation accuracy of 95%. And performs well on the test data. The segmented images are then used to train a lightweight CNN Architecture "MobileNet" using some pretrained weights from imagenet dataset and produced validation accuracy of 84.73% which is pretty good performance.