FFT-Assisted U-Net Architecture for Improved Skin Lesion Segmentation

Abstract

Dermatologists rely heavily on accurately segmenting skin lesions as it provides valuable insights into the clinical characteristics at the local and global levels. The diagnostic accuracy of correctly identifying skin lesions depends heavily on the quality of the segmentation. However, identifying clinical features from segmented images can be a tedious, subjective, and complex process due to the unique features and variations in the fine-grained appearance of skin lesion images. This study proposes a novel approach for skin lesion segmentation to address these challenges. By extracting Fast Fourier Transform (FFT) features from the skin lesion image and feeding them into U-net architecture while also feeding the original image into a separate U-net architecture, the results from both architectures are then concatenated to produce the final output. The developed method was trained and tested using the PH2 and ISIC-2018 datasets. The results demonstrated that combining features from various sources, such as the FFT and the original skin image, can assist in extracting deep features in various ways, resulting in a more discriminative and robust skin lesion segmentation approach. Additionally, the developed method achieved significantly better results in segmenting skin lesion images than state-of-the-art methods.