Transfer Learning-Based Encoder-Decoder Model for Skin Lesion Segmentation

Skin cancer is a highly dangerous form of cancer that affects numerous countries. Studies have shown that the timely detection of melanoma or skin cancer can lead to improved survival rates. Patients diagnosed with melanoma cancer at an early stage have a 90% chance of survival and tend to respond well to treatment. With this understanding, this study seeks to develop various architectures for the early detection of melanoma. Transfer learning, an approach that has garnered significant attention among researchers in solving computer vision problems, was employed in this project. In this study the U-Net architecture’s encoder was modified by replacine it with a pre-trained model to enhance its performance. The performance of the proposed segmentation model was evaluated using the ISIC-2018 dataset. The model recorded a dice coefficient score of 90.7% which is a 4.7% improvement on U-Net model (86%) for segmenting skin lesions. The model’s performance was further evaluated using other metrics such as recall (91.86%) and precision (91.13%). Subsequent analysis was conducted to determine the best hyper-parameters that provide the highest degree of performance when segmenting skin lesions. The results revealed that using the Efficient-Net pre-trained model as the encoder, PReLu activation function, and Tversky loss function yielded better performance in segmenting skin lesions.