Improving Loss Function for a Deep Neural Network for Lesion Segmentation

Identifying and segmenting lesions are challenging tasks in automatic analysis of endoscopic images in computer aided diagnosis systems. Models based on encoder-decoder architectures have been proposed to segment lesions with promising results. However, those approaches have limitations in modeling the local appearance, dealing with imbalanced data, and over-fitting. This paper proposes a novel method to address these limitations. We propose to improve a state-of-the-art encoder-decoder based model for image segmentation, named as Focal-Binary Cross Entropy (BCE)-Intersection over Union (IoU) loss for Feature Pyramid Network (FocalFPNet), by introducing a new loss function for training. The novel loss function is called Focal-BCE-IoU (FBI), consisting three terms, a Focal term, a BCE term, and an IoU term. In addition, we employ the Lasso regularization sparsity technique in learning to reduce over-fitting. As a result, our proposed model effectively segments lesions of various sizes and shapes, thereby improving the accuracy of the lesion segmentation task. Specifically, the Dice scores of the proposed model on the Esophageal cancer and the Peptic ulcer are 86.54 and 70.05, respectively. © 2023 ACM.