ADVANCED DOMAIN ADAPTATION FOR SKIN DISEASE SEGMENTATION AND CLASSIFICATION USING BOOTSTRAPPING OF FINE-TUNED DEEP LEARNER

Abstract

In medical diagnostic systems, the most challenging task is to segment and classify the varieties of skin disorders from dermoscopic images. For this purpose, Bootstrapping of Fine-tuned Segmentation and Classification Network (BF-SegClassNet) model was designed, which uses (i) cycle-Generative Adversarial Network (GAN) as domain adaptation, (ii) modified SegNet as segmentation and (iii) fine-tuned ResNet18 with Bootstrapping as classification. But, the efficiency of cycle-GAN was degraded if the source domain differs largely from the target domain. Hence, in this article, a Fuzzy Transfer Learning (FTL) model is developed based on fuzzy logic as domain adaptation. In this model, 2 different stages are performed such as training and adaptation. During the training stage, the source labeled data is used to build the Fuzzy Inference System (FIS), which extracts information from the source and transfers it to the target domain. The fuzzy sets and fuzzy rules created by an Adhoc Data-Driven Learning (ADDL) activity are included in the FIS. The created source FIS and the target data are used in the adaptation stage to adapt the fuzzy rule and the fuzzy rule base from the FIS to extract dissimilarities in the data and help bridge the contextual gap between the source and target. Thus, this FTL model is applied instead of cycleGAN to create more samples, which are further partitioned and classified by the BF-SegClassNet model efficiently. Finally, the testing outcomes exhibit that the FTL model attains a mean accuracy of 98.08% for the HAM dataset compared to the other GAN models.