A Modified UNet for Skin Lesion Segmentation using Transfer Learning

The advancement of deep learning techniques, such as Convolutional Neural Networks (CNN) provide mesmerizing results in dealing with medical image problems and can perform better with fewer data motivating researchers to design automatic systems which can diagnose diseases. Recent increases in the prevalence of skin cancer pose significant risks to people's health. As there can be low contrast between the lesion area and the lesion background, varieties of colors, noise artifacts, and irregular lesion boundaries, image segmentation is one of the most critical and complex steps in designing a computer-based diagnosis system. This work introduces an effective skin lesion segmentation method to overcome these challenges. Our proposed method replaces the UNet encoder with a pre-trained CNN model, uses the power of transfer learning by importing layers from the Xception, and uses the network's weights to calculate rich features. Fine-tuning on the Xception encoder improves the results and optimized training parameters improve the network's performance. Moreover, the decoder of UNet restores the image's spatial information. The proposed network is trained on a publicly available benchmarked dataset ISIC 2017 and tested on a completely different dataset PH2. The experimental results indicate that this proposed transfer learning model provides satisfactory results with segmentation accuracy of 93.39%, IoU of 83.43%, and Dice Coefficient of 90.56% on the ISIC 2017 dataset.