Research On Tissue Segmentation Algorithms For Endoscopic Esophageal OCT Images

Esophageal is a common site of inflammation and tumorigenesis.Because its early symptoms are unclear and difficult to detect,the diseases are usually found to be in the middle or late stage,which greatly increases the difficulty for follow-up treatment.Therefore,early diagnosis and intervention for esophageal diseases are necessary.Optical Coherence Tomography(OCT)which has the characteristics of non-invasive,real-time and high-resolution imaging is often used in the diagnosis of endoscopic diseases by combining with endoscopic technology.Research shows that the thickness of esophageal tissue is closely related to disease,but when in the face of a large number of collected images,manual annotation is time-consuming and labor-intensive.Therefore,it is important to design an automatic segmentation algorithm suitable for endoscopic esophageal OCT images.Under the condition with less experimental data,this paper designs a method of geodesic distance combined with superpixel for esophageal OCT image segmentation.First of all,a set of pre-processing process is proposed to solve the difficulty of image segmentation,which realizes noise suppression and interference removal,and lays a foundation for subsequent segmentation.Secondly,it is easy to fall into the local optimal problem for geodesic distance method.In the specific segmentation of the boundaries,this paper uses the coarse positioning first and then fine segmentation strategy to realize the boundary adaptively limited search area.Not only can the image with uniform layer thickness be segmented well,but also the image with large variation in layer thickness can be processed effectively.When the experimental data is sufficient,this paper uses deep learning methods to process and analyze the esophageal OCT images.Based on the U-Net model,two improved methods are proposed:a dual-stage U-Net framework and a SEN-Net model.The dual-stage U-Net framework first makes a binary classification,separates the effective tissue area from the background area,and then sends the mask made by classification result and the original image to the second stage for multi-classification to achieve the final segmentation effect.The pooled index module is added to the specific segmentation network,which effectively improves the binary classification result.This method can effectively avoid the problem of incomplete topology by forming two simple improved U-Net models into a dual-stage processing framework.However,the overall experimental process cannot fully realize end-to-end operation,a SEN-Net model is proposed.This model introduces three modules:pooled index,pyramid image input and SE-block,which together affect the final segmentation effect of the network.The above methods were tested on two healthy guinea pigs and one EoE-like model,and compared with other algorithms in all aspects.The experimental results show that when a method of geodetic distance combined with superpixel is used to segment the tissue layer,the absolute error for average layer thickness between the automatic results and the manual average labeling is a maximum of 1.25 ?m,and the maximum absolute error from the manual average boundary is 3.11 ?m.Compared with other algorithms,it is closer to manual labeling.The Dice similarity coefficient of the dual-stage U-Net framework is all more than 78.49%compared with manual labeling,which proves the effectiveness of this method.Through the ablation experiment,the introduction of the three modules in the SEN-Net model has a 1%?2%improvement in the Dice similarity coefficient compared with the original U-Net.For the EoE-like model,the maximum improvement can reach 5.98%.And finally achieve accurate segmentation for the endoscopic esophageal OCT image tissue layers.