Deep Learning With Convolutional Neural Network For Differentiation Of Renal Parenchymal Tumors At MRI

[Objective]To investigate the feasibility and diagnostic efficiency of convolutional neural network based on MRI T2W/FS sequence to differentiate renal parenchymal tumors.[Materials and Methods]The retrospective study included 186 patients with pathologically proven renal parenchymal tumors who undergone kidney MRI scans before surgery(74 in clear cell renal carcinoma,39 in papillary renal cell carcinoma,43 in chromophobe cell carcinoma and 30 in angiomyolipoma).The patients were divided into four groups according to the pathological findings.The T2-weighted fat saturation sequence images were copied from the PACS,and the region of interest(ROI)was manually segmented with the software named ITK-SNAP.Convolutional neural network(CNN)was used to investigate the ability of deep learning network for pathological classification of lesions.All the ROIs were divided into the size of 64*64.Finally,6501 image patches were used for training and 460 images patches were used for testing.Receiver operating characteristic(ROC)curve analysis was performed to estimate the performance of the CNN model,accuracy,precision,sensitivity,specificity,F1-score and area under the curve(AUC)were calculated.[Result]The experimental results demonstrated that the model has a 71%overall accuracy,a 86%average accuracy,with a macro-average AUC of 0.90.The AUCs for ccRCC,CRCC,AML,and PRCC were 0.92,0.88,0.93,0.87,respectively.[Conclusion]This study showed that the deep learning convolutional neural network based on MRI T2W/FS sequence was feasible to classify different types of renal parenchymal tumors and had a relatively high diagnostic efficiency.[Objective]To investigate the feasibility of convolutional neural network based on DCE-MRI sequence to differentiate renal parenchymal tumors.[Materials and Methods]The retrospective study included 149 patients with pathologically proven renal parenchymal tumors who undergone kidney DCE-MRI scans before surgery(71 in clear cell renal carcinoma,36 in chromophobe cell carcinoma,14 in papillary renal cell carcinoma,and 28 in angiomyolipoma).The patients were divided into four groups according to the postoperative pathological classification.The DCE-MRI images were copied from the PACS.Images of late corticomedullary phase,nephrographic phase and excretory phase were retained for image registration.Then the region of interest(ROI)in all three phases’ images was manually segmented with the software named ITK-SNAP one by one.Convolutional neural network(CNN)was used to investigate the usability of deep learning network for pathological classification of lesions.All the ROIs were divided into the size of 64*64*3.Finally,14066 image patches were used for training and 1626 image patches were used for testing.Accuracy,precision,sensitivity,specificity,F1-score and area under the curve(AUC)were calculated to estimate the performance of the CNN model.[Result]The results demonstrated that the model has a 75%overall accuracy,a 88%average accuracy,with a macro-average AUC of 0.90.The F1-scores for ccRCC,CRCC,pRCC and AML were 0.79,0.74,0.78,0.72,respectively.And the AUCs for ccRCC,CRCC,AML,and PRCC were 0.94,0.87,0.84,0.93,respectively.[Conclusion]This study demonstrated the feasibility of classifying renal parenchymal tumors by using artificial intelligence methods based on DCE-MRI sequence.