Application Of Multi-scale Neural Network In Identification Of Endometrial Precancerous Lesions

Endometrial intraepithelial neoplasia（EIN）has been proven to be a precancerous lesion of endometrial cancer（EC）,and accurate diagnosis of EIN is particularly important for the health of the female reproductive system.EIN is also known as atypical hyperplasia and is distinguished from benign hyperplasia without atypia（Hw A）and normal endometrium（NE）,but they are very similar in endometrial histomorphology and are prone to misdiagnosis and omission in clinical diagnosis,meanwhile the scarcity of pathologists is also a major status quo in the current Chinese pathology community.Computer-aided diagnosis relieves the diagnostic pressure of pathologists to a great extent,especially deep learning-based analysis of pathological images,which is expected to improve both diagnostic efficiency and accuracy.Considering that most previous studies have been limited to single-scale feature analysis on endometrial cell nuclei,this paper constructs deep learning models to identify endometrial lesions from single-scale and multi-scale perspectives,respectively.The specific research work is as follows:（1）Global Net,a single-scale network based on global image,and Local Net,a single-scale network based on multiple local images,are designed to perform the endometrial classification task（three classes:NE,Hw A and EIN）and the precancerous screening task（two classes:NE&Hw A and EIN）,respectively.Global Net uses fully supervised learning to extract global features on the global image,while Local Net deploys multi-instance learning based on weak supervision to extract local features on four sub-images.The accuracies of Global Net and Local Net on the endometrial classification task are 92.52%and 92.81%,respectively,and the area under the curves（AUCs）on the pre-cancer screening task is 0.9687 and 0.9668,respectively.Although the classification performance of Global Net and Local Net in both tasks are at the same level（p>0.05）,the Global Net is better at identifying Hw A,while Local Net is better at identifying NE.（2）The experience shows that both global（cytological changes relative to the background）and local features（gland-to-mesenchymal ratio and lesion size）contribute to the identification of endometrial lesions.Also,based on the complementary nature of Global Net and Local Net in identifying categories,we propose the G2LNet,a multi-scale network combining global and local features,for diagnosing endometrial tissue images.G2LNet extracts contextual features from the global image and deploys multi-instance learning to extract texture features from the multiple local images,and subsequently fuses the contextual and texture features to complete the endometrial classification and the pre-cancer screening.G2LNet achieved an accuracy and F1-score of 97.01%and 0.9698respectively in endometrial classification,and an accuracy and AUC of 99.07%and 0.9902in precancer screening,respectively.G2LNet outperformed the model based on single-scale image in both tasks（p<0.05）,including Global Net and Local Net,validating the effectiveness of the multi-scale network structure.G2LNet also performed significantly better than the state-of-the-art methods including HIENet and SVM_CNNin both tasks（p<0.05）,demonstrating G2LNet is better performance in the task of identifying endometrial lesions.（3）For the endometrial classification task on externally validated data,G2LNet achieve an accuracy and F1-score of 95.34%and 0.9531,respectively.The accuracy is comparable to that of a pathologist of moderate seniority（with 6 years of pathology experience）,but G2LNet is better at identifying EIN in the classification,while the physician is better at identifying NE and Hw A.In addition,we use the Grad＿Cam to visualize the regions of interest of G2LNet in classification,providing pathologists with more intuitive interpretation.The regions of interest show that some cases can be correctly classified precisely due to the combination of global and local features.