Research On Laryngeal Image Generation And Blood Vessel Classification Method Under Narrow-Band Imaging

As a common type of cancer in the head and neck region,early symptoms of laryngeal cancer are often not apparent.By the time it begins to affect the patient’s quality of life,the disease has often progressed to middle or late stages.Therefore,early detection of suspicious lesions is advantageous for timely diagnosis and determination of subsequent treatment options.Currently,the widely used technique for laryngeal examination is the endoscopic white light imaging technology,which has lower sensitivity for lesions occurring in the superficial layer of the mucosa.Narrow-band imaging(NBI)technology has become a powerful tool for detecting early cancer due to its ability to clearly display tiny lesions on the surface of the mucosa.However,this imaging technology depends on specific equipment and the operation of the examining physician.The subtle differences in visual perception make it difficult for inexperienced physicians to identify the type of blood vessels on the mucosal surface,resulting in the omission of early lesions.Therefore,the development of an automated early cancer detection system has important practical significance.Existing automated image analysis methods mostly use fully supervised learning,which requires a large amount of annotated data.Considering the challenge of medical image collection and labelling,this paper adopts unsupervised and self-supervised learning approaches that have demonstrated greater potential in the study of this type of problem.The following research work in the field of laryngeal image generation and classification has been carried out to address the high dependence of existing examination processes on equipment and manual handling.(1)In response to the dependence on endoscopic equipment and subjective judgments of operators during laryngoscopy,an unsupervised method for generating laryngeal endoscopic images is proposed.This method does not require fully matched image pairs and can generate narrow band images corresponding to white light images,highlighting the morphology of the superficial microvasculature of the mucosa while preserving the complex structure features of the larynx.This allows physicians to quickly and accurately detect lesions.The proposed method optimizes the generator network structure by embedding a multi-scale cross-layer adaptive module to overcome the limitations of unsupervised image translation methods based on cycle consistency in capturing local image details.In addition,a structure difference loss is designed to establish a connection between the input image and the generated image by minimizing the structure difference between the two images,thereby improving the realism and completeness of the generated images.Furthermore,a feature denoising module is introduced to reduce the interference of noise information and improve the quality of generated images.Through ablation experiments and comparative experiments,the proposed method is validated,and the generated images are of good quality with relatively stable performance.This is supported by quantitative analysis results,as well as image quality and local detail analysis.(2)Given the irregular morphology and lack of accurate labelling of superficial blood vessels in laryngeal mucosa,combined with the characteristics of local blood vessels in the larynx,a two-stage classification method of laryngeal blood vessels based on self-supervised learning is proposed.In the first stage,the patch-level feature extraction network is trained using the contrast learning framework Sim CLR to learn the intrinsic features of unlabelled data.Then,the label information of patch is used for supervised training to learn the difference characteristics between classes.In the second stage,SACO-Net shares the feature extraction network with the first stage,aggregates the local features of the larynx image and establishes the long-distance dependency between patches through self-attention mechanism to provide the global information for the backbone.After the prediction results are output by the network,the category activation distillation branch is designed to use the local feature classification network knowledge from stage 1 as an auxiliary supervisory signal to calculate the class offset loss,prompting the network to focus on key regions and improve vessel classification performance.In addition to the ablation experiment designed for the method itself,this paper also makes a quantitative comparison with other mainstream supervised learning methods,and further verifies the performance and potential of the proposed method by combining the visualization results of the class activation map and the analysis of sample incremental experiments.