Development And Validation Of Cervical Cancer Lymph Node Metastasis Prediction Model Based On Deep Learning And Computational Pathology

BACKGROUND:Cervical cancer is one of the important public health problems threatening women’s life and health.The incidence of cervical cancer tends to be younger,and lymph node metastasis is one of the important factors affecting the prognosis of patients with cervical cancer.Currently,the main treatment options for early cervical cancer are extensive hysterectomy and pelvic lymph node dissection and lymph node dissection is used to assess the presence of lymph node metastasis in patients.However,there are many complications after lymph node dissection,which has a great impact on the quality of life of patients.Preoperative identification of patients with very low probability of lymph node metastasis,no lymph node dissection or sentinel lymph node biopsy,can improve patients’ quality of life.For patients with very high risk of lymph node metastasis,radical chemoradiotherapy can be the first choice.Therefore,it is very important to accurately evaluate the status of lymph node metastasis in patients before surgery,which can provide reference for the formulation of individualized treatment for patients with cervical cancer.OBJECTIVE:This study aims to use deep learning model to extract pathological features based on H&E staining pathological sections of primary tumor tissue,and then predict lymph node metastasis of patients,to provide reference for individualized treatment.METHODS:In this study,277 cases of Qilu Hospital of Shandong University and 168 cases of operable early cervical cancer patients from other hospitals(Second Hospital of Shandong University,Shandong Provincial Hospital,Affiliated Hospital of Jining Medical College,Jinan First People’s Hospital,Taian Central Hospital)were retrospectively analyzed.The clinicopathological data and postoperative H&E staining pathological sections were analyzed.The pathologist selected 1-5 H&E staining pathological sections for each case and digitized them by scanner.A total of 1068 WSIs were obtained.The postoperative H&E staining pathological sections WSI of cervical cancer treated in Qilu Hospital was used as the training cohort(277 cases,WSI=868),and the postoperative H&E staining pathological sections WSI of patients from other hospitals was used as the external validation cohort(168 cases,WSI=200).The deep learning model used the vision transformer(ViT)and recurrent neural network(RNN)networks as frameworks.In order to further migrate the application of the model to preoperative biopsy pathology,95 patients with operable early primary cervical cancer diagnosed in Qilu Hospital of Shandong University from May 2008 to March 2021 were included,and 108 H&E staining pathological slices of preoperative biopsy were retrieved and reviewed(migration test cohort,WSI=108).Transfer learning test is carried out on deep learning model.In this study,to further validate the performance of the deep learning model,the accuracy of imaging examinations such as MRI or CT for lymph node metastasis evaluation in both the training set and migration test set was calculated.This study further analyzed the relationship between clinicopathological indicators and lymph node metastasis,while constructing a prediction model of clinicopathological factors to compare the performance of deep learning models with preoperative imaging examinations and clinicopathological models in assessing lymph node metastasis.RESULTS:In the training cohort,the area under the operating characteristic curve(AUROC)of the deep learning model is 90.3%,the sensitivity and specificity are 84.0%and 96.7%,respectively,and the F1_score is 90.7%.In the external validation cohort,the deep learning model AUROC is 85.0%,sensitivity and specificity are 73.8%and 96.3%,respectively,and F1_score is 86.5%.In the migration test cohort,the AUROC is 88.2%,the sensitivity and specificity are 81.6%and 94.8%,respectively,and the F1_score is 88.8%.In the training cohort,the sensitivity of imaging examination to determine lymph node metastasis is 52.8%,the specificity is 78.9%,and the AUROC is 65.9%.In the migration test cohort,the sensitivity of imaging examination to determine lymph node metastasis is 47.4%,the specificity is 91.7%,and the AUROC value is 69.5%.Clinicopathological indicators were analyzed,and LVI,tumor diameter and parametrial infiltration were independent risk factors for lymph node metastasis.The above clinicopathological indicators were included in the construction of the clinicopathological model.In the training cohort,the sensitivity of the clinicopathological model is 72.7%,the specificity is 81.5%,and the AUROC is 83.5%;in the external validation cohort,the sensitivity of the clinicopathological model is 69.8%,the specificity is 76.0%,and the AUROC is 76.3%;in the migration test cohort,the sensitivity of the clinicopathological model is 49.3%,the specificity is 100.0%and the AUROC value is 82.2%.Despite the inclusion of some postoperative pathological parameters in the clinicopathology model,the accuracy of the deep learning model was still higher than the clinicopathology model in the migration test cohort.CONCLUSION:The accuracy of the model based on deep learning and computational pathology was higher than the clinicopathological model and imaging examination in the evaluation of lymph node status in patients with cervical cancer.Deep learning and computational pathology model can be used to predict the status of lymph nodes in patients with early cervical cancer.Preoperative cervical biopsy H&E staining pathological section WSI can be used to predict the status of lymph nodes in patients with cervical cancer,to formulate individualized treatment plans for patients with cervical cancer.