A Study on Multi-Class Laryngoscope Image Semantic Segmentation Using Attention U-Net and a Composite Loss

Laryngoscopy is an important clinical diagnostic tool in otolaryngology. Precisely segmenting multiple types of structures in laryngoscopic images can aid physicians in making informed decisions. However, traditional U-Net models struggle with complex multi-class structures and class imbalance in such images. We propose a multi-class semantic segmentation approach based on the Attention U-Net architecture, trained with a composite loss function that combines cross-entropy and a multi-class Dice loss. Experimental results show that compared to conventional U-Net, the proposed method significantly improves multi-class segmentation accuracy and generalization.