Magnetic Resonance Imaging-based Artificial Intelligence-aided Diagnosis Of Obstetrics And Gynecology Diseases

Among obstetrical and gynecological conditions,the prognosis of ovarian cancer and the diagnosis of placenta accreta spectrum are emerging scenarios for magnetic resonance imaging(MRI)applications.Compared to ultrasound and CT,the most common examinations in obstetrics and gynecology,MRI is uniquely suited for ovarian and fetal examinations due to its excellent soft tissue contrast and lack of radioactivity.In the past,due to the low detection rate of MRI for good or bad prognosis and the high dependence on clinician expertise for the determination of placental tissue invasion,its effective information was not fully utilized.In recent years,with the continuous development of informatics,great progress has been made in the application of Artificial Intelligence(AI)in medical imaging,which can help radiologists automatically discover and identify lesions,assess disease progression,and develop treatment plans in medical imaging.Therefore,in this paper,we explored the information in MRI using AI to deeply explore the auxiliary diagnosis of ovarian cancer and placenta accreta spectrum.(1)Prognosis of epithelial ovarian cancer: The postoperative survival status of ovarian cancer is influenced by several factors,and accurate prognosis helps to design treatment plans.We utilized radiomics to establish a model to predict the prognostic status,combined the clinical indicators with the radiomics features from T2WI(T2-weighted imaging),T1WI(T1-weighted imaging),T1CE(contrast-enhanced T1-weighted imaging)and DWI(diffusion weighted imaging).The results demonstrated that the combined MR-based radimics-clinical model has high accuracy for prognosis and helps clinical assessment.(2)Diagnosis of antenatal placenta accreta spectrum: Antenatal diagnosis of antenatal placenta accreta spectrum can help prevent maternal and fetal risks.And accurate diagnosis relies heavily on radiologist’ experience.To reduce radiologist’workload,we performed automatic placenta segmentation on T2 WI using nn U-Net.Then we combined the segmentation with radiomics methods to extract key features from the placenta and combined with clinical features to establish a combined radiomics-clinical model.The results showed that the combined model based on prenatal T2 WI has high accuracy in the diagnosis of placenta accreta spectrum,which can provide effective help for local hospitals as well as young doctors.(3)Fully automated diagnosis of placenta accreta spectrum: The diagnostic radiomics model for placenta accreta spectrum requires a radiologist to read the MRI and provide a clinical diagnosis of the placenta position,which severely limits the predictive scenario of the diagnostic model.In order to fully automate the disease diagnosis,we combined deep learning methods for placenta segmentation,position detection,and placenta accreta spectrum diagnosis with a specially designed image resampling method,respectively.The results show that the deep learning model can perform placenta accreta spectrum diagnosis fully automatically.Our designed resampling method makes the network focus on the utero-placental borderline where the lesion may occur,which can effectively improve the performance of the model and obtain a competitive model with less training data.A fully automated predictive model can be an effective cue for clinicians’ decision making.This work uses AI to achieve(1)the prognosis of epithelial ovarian cancer using the ovarian region outlined by the radiologists;(2)automatic segmentation of the placental region in the MRI,combined with the information of the placental location marked by the radiologists for placenta accreta spectrum diagnosis;and(3)fully automated segmentation of the placental region,prediction of the placental location,and combination of the resampling algorithm for placenta accreta spectrum diagnosis,which can effectively provide an aid for clinicians.