Skin Lesion Detection And Segmentation Using Deep Transfer Learning Approach

Abstract

The recent depletion of the ozone layer due to industrial pollution has resulted in an increase in UV radiation, which is a significant environmental risk factor for invasive skin cancer and other keratinocyte tumors. Over the last few decades, cancerous deaths have increased alarmingly throughout the world. For dermatological diagnosis, deep learning has been practised successfully. This study, therefore, presents a deep-learning-based technique for automatically segmenting skin lesions and identifying skin cancers from dermoscopy images. The lesion is segmented from the adjoining areas of the skin by using the U-Net which limits the use of deep neural networks. This problem is resolved through the techniques of data augmentation and transfer learning. In our studies, we added a variety of augmentation effects to the training images to improve the data samples, and we utilised U-Net with dropout to solve the overfitting issue. On two separate datasets, the model was analysed. On the ISIC 2018 dataset, it had a mean Jaccard Index of 0.80 and an average dice score of 0.87. Using a transfer learning strategy, the trained model was evaluated against the PH dataset and got a mean dice score of 0.93 and an average Jaccard index of 0.87. A DCNN-SVM model was used to classify malignant melanoma, and to evaluate how well transfer learning is being used in the field of dermatological diagnosis, we compared cutting-edge deep architectures as feature extractors. On the PH2 dataset, our top model had an average accuracy of 93%.