A Preliminary Study On Semantic Segmentation Algorithm In The Construction Of Benign And Malignant Pathological Diagnosis Model Of Gastric Tissue

Objective: This study aimed to establish a fully supervised deep-learning model based on semantic segmentation algorithm through the segmentation and training of pathology images of benign and malignant gastric lesions under hematoxylin–eosin(HE)staining,as well as a pathology-aided diagnosis model of benign and malignant gastric lesions.Methods:(1)Pathological sections of benign and malignant gastric tissues stained with HE were collected from the Department of Pathology of the First Affiliated Hospital of Medical College of Shihezi University.It included 200 cases(33.3%)of gastritis,160 cases(26.6%)of low-grade gastric intraepithelial neoplasia(LGIN),40 cases(6.7%)of high-grade gastric intraepithelial neoplasia(HGIN)mucosal biopsy HE-stained pathological sections,and 200 cases(33.3%)of gastric cancer(GC)mucosal biopsy HE-stained pathological sections.(2)The digital pathology images of gastric tissue were scanned with a digital pathological-section scanner(one high-quality image for each case,including 115 gastritis,80 LGIN,21 HGIN,and 114 GC)as a training set to mark the outline of gastric tissue.Eleven pathological tissue tags were formed,including gastritis–gland,intestinal metaplasia–gland,LGIN–gland,HGIN–gland,mucous muscle,lymphoid tissue,gastritis–mucosal epithelium,LGIN–mucosal epithelium,HGIN–mucosal epithelium,cancer area,stroma,and others).Moreover,digital pathology images of benign and malignant gastric tissues were selected as the test set(one high-quality image was selected for each case),including40 gastritis,30 LGIN,10 HGIN,and 40 GC.(3)The labeled glands and mucosal epithelium labels were selected to extract the morphological parameters,and SPSS software was used for statistical analysis.(4)Deeplabe-V3,PSPNet,and HRNet network-model algorithms were used for the preprocessing and semantic segmentation of benign and malignant gastric tissue images in the training set.(5)According to the semantic-segmentation index,the optimal network-model algorithm was selected and combined with the image-recognition strategy.First,the 120 images in the test set were used for the classification and recognition of GC and non-GC,and the GC image was selected;second,the non-GC image was classified and recognized for gastritis and intraepithelial neoplasia,and the image of intraepithelial neoplasia was screened out;finally,the screened image of intraepithelial neoplasia was classified and recognized by LGIN and HGIN,and the HGIN image was screened out.Results:(1)A total of 62782 tags of different tissue types in benign and malignant gastric lesions were established,including 38459 gastritis–glands,5511 LGIN–glands,512 HGIN–glands,1146 intestinal metaplastia–glands,71 mucosal muscles,50 lymphoid tissues,4413 gastritis–mucosal epithelium,2635LGIN–mucosal epithelium,252 HGIN–mucosal epithelium,5379 cancerous areas,and 4354 stroma and others.(2)The analysis and statistics of morphological parameters showed significant statistical differences in the morphological characteristics of glands and mucosal epithelium of gastritis,LGIN,and HGIN.These parameters included polarity,eccentricity,the total number of pixels in the convex hull,the total number of pixels filled between the region and the outer frame,the average intensity in the region,the diameter of the circle that was the same as the area of the region,the area of the region,and the ratio of the area of the boundary(P<0.05).(3)According to the image-recognition strategy,the accuracy,precision,sensitivity,and specificity of PSPNet network-model algorithm for GC and non-GC classification in all images were87.5%,72.7%,100%,and 81.3%,respectively.Meanwhile,the accuracy,precision,sensitivity,and specificity of HRNet network-model algorithm for gastritis and intraepithelial neoplasia were 86.25%,100%,72.5%,and 100%,respectively.Finally,the accuracy,precision,sensitivity,and specificity of HRNet network-model algorithm for LGIN and HGIN image classification were 65%,37.5%,60%,and 66.7%,respectively.Conclusion:(1)Parameter features extracted by semantic-segmentation algorithm can be used as morphological indices to distinguish gastritis,LGIN and HGIN glands,and mucosal epithelium.(2)PSPNet and HRNet network-model algorithm combined with image-recognition strategy can preliminarily classify and identify the lesion areas of benign and malignant gastric lesions.