Automatic diagnosis of laryngeal diseases and segmentation of laryngeal cancer based on deep learning technology

Abstract Background Laryngeal cancer is currently a common clinical malignancy, and its early diagnosis is of great significance for improving patient survival rates. With the widespread application of deep learning technology in the field of medical image analysis, this study aims to establish an intelligent diagnostic system for early screening and precise diagnosis of laryngeal cancer. Methods We design the ScSE and GC modules and develop a lightweight laryngeal disease classification network named LLLDNet, employing a combined enhancement strategy of transfer learning and knowledge distillation to improve the accuracy of laryngeal disease image classification. we also design the HA and MSFF modules to develop a laryngeal cancer segmentation network named HAM-Unet, which integrates hybrid attention and multi-scale feature fusion. Finally, we conduct experiments on the proposed LLLDNet classification network and HAM-Unet segmentation network using datasets of laryngeal diseases and laryngeal cancer, respectively. Results Experimental results demonstrate that LLLDNet achieves an accuracy of 93.0% and a recall of 92.4%. While improving various metrics including accuracy, LLLDNet reduces the number of model parameters to only 0.425M and FLOPs to 294M, achieving sufficient lightweight design, conserving computational resources, and accelerating inference. The HAM-Unet network attains an IOU of 90.86% and an accuracy of 93.92%, significantly outperforming other segmentation models and enabling more precise delineation of laryngeal cancer lesion regions.