Application Of Artificial Intelligence In Cystoscopy Image Diagnosis

Objective To establish a deep learning model in cystoscopy images based on semantic segmentation algorithm,classify cystoscopy images of bladder tumors,distinguish pathological grades,and provide basis for the treatment of bladder tumors and prognostic follow-up.Methods Cystoscopy images of 261 patients with bladder tumor admitted to our hospital from January 2010 to August 2021 were retrospectively collected.According to the characteristics of cystoscopic images of bladder tumors with different pathological grades,the images were divided into 3 categories,including 116 cases of papilloma or papilloma with low malignant potential,185 images.63 cases with lowgrade urothelial carcinoma,97 images;42 cases of high-grade urothelial carcinoma,62 images.After data amplification,2408 images were obtained and randomly divided at a ratio of 8:1:1 to get the training set,test set and verification set.The multilevel feature semantic fusion network(MFSFNet)was used to segment the image.It was compared with four classical semantic segmentation models on our data set.Finally,external data sets were grouped according to clinical information(whether there is infiltration of muscle layer,gender,age,smoking,and family history of bladder cancer)to explore the clinical application value of the model and predict the classification of cystoscopy pictures.Mean Pixel Accuracy(MPA),Mean Intersection over Union,mean pixel accuracy(mpa),mean intersection over Union,MIoU),cross entropy loss function,model Parameters and Frame Per Second(FPS)were used as evaluation indexes of model experiment segmentation,and the accuracy of model image classification was calculated.Results After image recognition training,MFSFNet distinguished papilloma or low-malignant potential papilloma,low-grade urothelial carcinoma,and high-grade urothelial carcinoma with 95.93% MPA and 92.54% MIoU,and its image classification accuracy was 96%,which had certain advantages compared with the other 4 types of classical models.In the grouping experiment,the overall image classification accuracy was 51%,and the non-invasive group,male group,<60 years old group,smoking group and the group with a family history of bladder cancer had higher accuracy,which were 56%,56%,54%,53% and 71%,respectively.Conclusion MFSFNet network model has a good effect on the segmentation of bladder tumors on cystoscopy images,and can better classify bladder tumor images,which proves the application prospect of deep learning semantic segmentation model in the task of bladder tumor segmentation and has certain clinical application value.It can provide valuable information for early diagnosis classification and personalized treatment of patients with bladder tumor,and promote the development of human health.