P461 Deep learning model for endoscopic differential diagnosis between intestinal lymphoma and Crohn's disease: a multicenter study

Abstract

Abstract Background Primary and secondary intestinal lymphoma (IL) are a relatively rare cancer, while easily misdiagnosed due to the unclear aetiology and unspecific clinical symptoms. Moreover, the endoscopic features of IL vary depending on pathological types, making it difficult to distinguish from Crohn's disease (CD). This study aims to develop deep learning models based on endoscopic images, facilitating the differential diagnosis of IL and CD. Methods We retrospectively recruited 271 patients with IL and 232 patients with CD from 6 tertiary centres in China. With out-of-focus or poor-bowel-prepared images excluded, we used 1739 IL and 1649 CD images for internal and external validation to build deep learning models. For internal validation, we used five-fold cross-validation with 1517 IL images and 1428 CD images from 5 centres randomly assigned without any patient overlap. 222 IL and 221 CD images from the 6th centre were used for independent external validation to measure the generalization ability of models. Further, we compared the performance of models with that of clinical physicians consisting of 2 junior physicians and 2 endoscopy specialists with more than ten years of clinical experience. We evaluated models using metrics such as accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) and used the kappa statistic to measure the inter-rater reliability between models and clinical physicians. Results We developed multiple convolutional neural network models, including ResNet18, ResNet34, ResNet50, VGG16, VGG19, DenseNet121, Inception v3, and Xception. The top 3 models in the internal validation were Inception v3, Xception, and DenseNet121. Inception v3 showed the best performance (accuracy: 83.1%; AUC: 0.910), while Xception performed better on average (average accuracy: 76.8%; average AUC: 0.834). In the external validation, the top 3 models were Inception v3, DenseNet121, and VGG19. Inception v3 achieved the highest accuracy of 81.9%, while DenseNet121 had the best AUC of 0.903. Compared with clinical physicians, Inception v3 performed better than junior physicians (average accuracy: 67.2%) and showed good agreement with the best endoscopy specialist (kappa: 0.604, p<0.05). Conclusion Convolutional neural network models using endoscopic images were able to differentiate PIL from CD, with better-diagnosing performance than junior physicians, good agreement with specialists, and good generation ability, which were expected to assist clinical physicians lacking experience to differentiate PIL and CD more precisely under endoscope.