基于深度学习的前列腺mpMRI序列分类的可行性研究

Feasibility of Prostate mpMRI Sequence Classification Based on Deep Learning

目的探讨深度学习在前列腺多参数磁共振成像(mpMRI)序列分类中的应用价值。方法回顾性搜集2012年11月至2021年6月2085个连续前列腺mpMRI检查,共41427个序列。由一位高级职称影像专家逐一阅图,将图像分为以下9个序列类别:表观扩散系数(ADC)、扩散加权成像(DWI)_High、DWI_Low、CE_Ce、CE_Pre、T_(1)WI_In、T_(1)WI_Opp、T_(2)WI_Fs和T_(2)WI_Nan。将41427个序列数据按照8∶1∶1的比例随机分为训练集(n=32988)、调优集(n=4235)和测试集(n=4204)用于训练深度学习分类网络。利用混淆矩阵对分类效能进行评价。结果在总体分类评价水平上,测试集中宏准确率(Macro accuracy)、宏F1(macro-F1)和微F1(micro-F1)分别为0.997、0.890和0.989。单一序列准确率为0.996~1.000,F1为0.983~1.000,召回率为0.989~1.000。结论深度学习网络在前列腺mpMRI序列分类中具有很高的准确率,可用于后续前列腺癌自动诊断模型的序列甄别。

Objective To explore the ability of a deep learning algorithm to classify the sequences of prostate multiparametric magnetic resonance imaging(mpMRI).Methods Between November 2012 and June 2021,a total of 2085 consecutive prostate mpMRI examinations with 41427 sequences were acquired from picture archiving and communication systems(PACS).A senior radiologist observed the image features and divided the images into the following nine categories:ADC,DWI_High,DWI_Low,CE_Ce,CE_Pre,T_(1)WI_In,T_(1)WI_Opp,T_(2)WI_Fs,and T_(2)WI_Nan.The images were randomly split into the datasets with an 8∶1∶1 ratio for training(32988 sequences),validation(4235 sequences),and testing(4204 sequences).A modified Med 3D network was trained to classify the images of prostate mpMRI.The classification efficiency was evaluated with the confusion matrix.Results For the overall evaluation of the classification,the Macro accuracy,Macro F1,and Micro F1 of the model were 0.997,0.890,and 0.989,respectively.For the classification of each single sequence,the accuracy of the model was 0.996 to 1.000,the F1 was 0.983 to 1.000,and the recall was 0.989 to 1.000.Conclusion The deep learning network could be used to classify the different image types with very high accuracy for prostate multiparametric MR examination.