A Unified Approach for Skin Lesion Classification on Generated Segmented Dataset

The development of deep learning techniques, such as Convolutional Neural Networks (CNN), has yielded mesmerizing results in the context of medical image problems. These techniques are capable of performing better with less data, which has encouraged researchers to develop automatic systems capable of diagnosing diseases. Skin cancer’s prevalence has substantially increased in recent years, which poses substantial health hazards. Skin image classification is one of the most critical and complex stages in the design of a computer-based diagnosis system due to the presence of low contrast between the lesion area and the lesion background, a variety of colors, noise artifacts, and irregular lesion boundaries. This work introduces a skin lesion classification system that is effective in classifying skin lesions as either benign or malignant. In our proposed method, the segmented dataset from the ISIC 2016 dataset is generated using a UNet architecture having a pre-trained Xception encoder. Subsequently, the dataset is employed to classify two classes using multiple pre-trained models. The experimental results suggest that the proposed transfer learning model yields satisfactory results, with a classification accuracy of 86.55%. This enhanced accuracy is obtained by resolving the data imbalance issue and augmenting the dataset with multiple augmentation techniques, resulting in an increase in accuracy from 81.01% to 86.55%.