Application Of Cascade Network Based On Deep Learning And Radiomics In Automatic Recognition Of Adrenal Incidentalomas

Background:The high prevalence of adrenal incidentalomas has become a public problem,but it is easy to be ignored due to its small size and low clinical attention.Objective:To develop a cascade network based on deep learning automatic segmentation algorithm and radiomics model for automatic recognition of adrenal incidentalomas on CT,and to discuss the clinical application value of the cascade network for automatic recognition of adrenal incidentalomas in the data of multiple centers.Methods:This retrospective,multicenter study included two cohorts(development cohort and test cohort).The development cohort included the imaging and clinical data of 443 patients who underwent adrenal CT scans in our center from January 2018 to August 2021.The development cohort was divided into training set and validation set according to 7:3 random stratified sampling.The imaging data and clinical data of 335 patients who underwent adrenal CT scan from January 2017 to January 2022 in three centers were included in the test cohort.The development cohort was manually delineated and labeled by three urologists and used to train the 3D Res U-Net and 3D U-Net for automatic segmentation of bilateral adrenal glands and the radiomics model for determining the presence of unilateral adrenal incidentalomas.In the development cohort,Dice similarity coefficient(DSC),Intersection over Union(IoU),95%quantile of Hausdorff surface distance(HD95)and Relative volume error(RVE)were used to evaluate the performance of the automatic adrenal segmentation model.Receiver operating characteristic curve(ROC),area under ROC curve(AUC),accuracy,sensitivity and specificity were used to evaluate the classification performance of the radiomics model.After training,the 3D Res U-Net model was connected with the radiomics model to form an automatic adrenal incidentalomas recognition cascade network.The test cohort was used for the external test of the automatic recognition cascade network for adrenal incidentalomas.The receiver operating characteristic curve(ROC),area under the ROC curve(AUC),accuracy,sensitivity and specificity were used to evaluate the performance of the cascade network in recognition adrenal incidentalomas.Results:The DSC,IoU,HD95 and RVE of the 3D Res U-Net segmentation model in the validation set were 85.94±5.74%,75.75±8.19%,3.17±3.13mm,7.20±8.75%,respectively,while the corresponding value of 3D U-Net in the validation set was 82.42±8.14%.70.82±10.53%,4.59±5.14 mm,12.93±13.06%.All segmentation metrics of 3D Res U-Net were significantly better than those of 3D U-Net(P<0.001).Based on the automatic segmentation images,24 and 6 imaging features of left and right adrenal incidentalomas were extracted and screened,respectively.The AUC of the random forest models established according to these imaging features in the validation set was 88.15%(95%CI 82.34-93.98%,left)and 87.90%(95%CI 78.49-97.31%,right).When the DSC was higher than 82.28%,there was no significant difference in the performance between the radiomics models based on automatic segmentation and manual segmentation.The AUC of the automatic adrenal recognition cascade network constructed by combining 3D Res U-Net model and random forest model was higher than 80%in the three test sets of the test cohort.Conclusion:The 3D Res U-Net automatic segmentation model based on deep learning can accurately segment the bilateral adrenal glands on CT images,and the subsequently established radiomics model can effectively detect adrenal incidentalomas.The cascade network based on automatic segmentation model and radiomics model can robustly and automatically recognition adrenal incidentalomas without manual delineation,which provides a new method for early recognition of adrenal incidentalomas.