A Research On Assisted Screening Of Early Esophageal Squamous Cell Carcinoma Based On Deep Learning

Esophageal cancer(EC)is a common type of gastrointestinal cancer,with a high morbidity and mortality.Esophageal squamous cell carcinoma(ESCC)is the predominant histological subtype of EC in China.Gastrointestinal endoscopy is currently the major technique used for the screening of early-stage ESCC.Due to variations and visual similarities of lesions in shapes,colors and textures in endoscopy,the efficiency and accuracy of diagnosing ESCC are significantly dependent on the clinical proficiency and subjective judgment of clinicians.Early-stage lesions are often missed under the white-light imaging endoscopy.Our work aims to develop and validate an artificial intelligence(AI)system based on convolutional neural network(CNN)for image classification and lesion localization of early-stage ESCC,to assist endoscopists in diagnosing early-stage ESCC under white-light imaging(WLI)endoscopy.In our study,a total of 10421 WLI images from 265,1 patients were collected from eleven hospitals.The endoscopic images from two main hospitals were used as the training set and internal validation set.The endoscopic images from the other nine hospitals were used as the external validation set.The main research contents in this thesis are as follows:(1)The classification of endoscopic images containing early-stage ESCC.According to the characteristics of esophageal endoscopic images,a CNN network architecture integrating the global attention mechanism and bilinear pooling was proposed for classifying esophageal lesions in endoscopic images.The ResNet50 was employed as the backbone network.The newly designed global channel attention module was adopted to recalibrate channel-wise feature responses.The bilinear pooling was used to enable the inter-layer interaction and cross-layer combination of features for improving the representation.For the internal validation set,the AUC value of our CNN model reached 0.993.It achieved an accuracy of 91.75%and a sensitivity of 96.64%at the image level.The accuracy and sensitivity were 88.38%and 90.17%,respectively,at the patient level.The proposed CNN model outperformed other state-of-the-art methods.In addition,the results in our external validation set demonstrated the good generalization ability of our method.(2)Comparison of resluts between the AI-assisted diagnostic system and the endoscopists.To evaluate the effectiveness of AI-assisted diagnostic system,the classification performance between AI-assisted approach and clinical endoscopists were compared.Eight endoscopists of different skill levels participated in our evaluation.Both the overall results and detailed results under different cancer conditions were analyzed.Moreover,the diagnostic performance of endoscopists with and without the AI assistance were also compared.The AI-assisted system achieved a high sensitivity comparable to endoscopists in the senior group(99.23%vs 96.41%,P=0.099)and better than endoscopists in the non-senior group(99.23%vs 89.62%/90.00%,P=0.001).The diagnostic ability of endoscopists was significantly improved with the AI assistance,in terms of the accuracy(75.12%vs 84.95%,P=0.008),specificity(63.29%vs 76.59%,P=0.017),and positive predictive value(64.95%vs 75.23%,P=0.006).(3)Lesion localization of early-stage ESCC.To localize lesions in cancer images,the patch-based classification strategy was utilized to generate the probability heat map related to early-stage ESCC.Image patches were randomly selected from cancer images in the training set to train the CNN model.The type of each image patch was determined by the delineations of endoscopists.During the evaluation,the full image was equidistantly sampled into a set of overlapping patches and inputted into the model.The probability heat map was compiled using the cancer probability value provided by the CNN model.The suspicious lesion area was obtained using a threshold value of 0.5.For the internal and external validation sets,the Dice coefficients that we obtained were 0.74 and 0.72,respectively.And the recalls obtained were 86.79%and 84.23%,respectively.The results demonstrated that this localization strategy could effectively localize the suspicious lesions.In conclusion,the AI-assisted ESCC diagnosis system proposed in this thesis showed an excellent performance for the classification of early-satge ESCC in WLI images,with a performance comparable to senior-level endoscopists.It could also effectively localize the suspicious lesions and provide their suspicious converage area for the endoscopists.The diagnostic ability of endoscopists could be significantly improved with the assistance of our AI-based approach and prototype system.Such an AI-assisted system could be used to reduce the dependence of the ESCC diagnosis on the experience and subjective judgment of endoscopists.In particular,it had a high potential to assist inexperienced endoscopists in diagnosing early-stage ESCC during WLI endoscopic examination procedures by improving both the efficient and accuracy.