MRI Radiological Features-based Deep Learning Predicts Lymphovascular Invasion in Early Cervical Cancer

ZOU Mingyang; LAN Bowen; LIAO Junjie; GUAN Qianwen

doi:10.3969/j.issn.1672-7185.2023.11.012

Abstract：Objective To explore the prediction of lymphatic vascular infiltration in early cervical cancer by MRI radiological characteristics based deep learning.Methods From January 2018 to December 2021,a total of 118 patients with early cervical cancer admitted to Huizhou Central People's Hospital were selected as research subjects,and were divided into 82 cases in the observation group(training set)and 36 cases in the control group(verification set)according to simple sampling.By collecting the general data,clinical data and imaging data of patients,two experienced doctors recorded the imaging data of cervical cancer patients to construct a clinical model,and then used ITK-SNAP software to outline the region of interest in the image,build a deep learning model by processing the data,and visualize the model with a Nomogram to evaluate the omics analysis of cervical cancer in predicting tumor grade and lymphatic invasion.The evaluation models such as clinical decision curve,discrimination and calibration were used.Finally,the model of the observation group was verified in the control group,and the receiver operating characteristic(ROC)curve was drawn to analyze the prediction performance.Results The six selected features have the smallest correlation and the strongest prediction ability.The area under the curve(AUC)were all greater than 0.60,indicating that lymphovascular space invasion(LVSI)of cervical cancer had a good predictive ability.The closer the 45° oblique line is to the origin,the better the prediction effect of the model is.The AUC values of the ROC curve were 0.815 and 0.778,respectively.When the threshold probability ranges from 0.07 to 0.77,the patients with early cervical cancer benefit well.Conclusion The model established in this study can effectively predict the lymphovascular space invasion of cervical cancer,and the diagnostic efficiency of the model is high,but the foreground data shows that the model construction is relatively rough,and it is still necessary to increase the sample size to further improve the model to better establish the prediction model.