人工智能深度学习对前列腺多序列MR图像分类的可行性研究

Deep learning for classification of multi-sequence MR images of the prostate

目的开发一种能够自动分辨前列腺多序列MR图像的人工智能(AI)工具。方法回顾性分析2017年5月至2018年12月华中科技大学同济医学院附属同济医院前列腺多序列MR图像。前列腺多序列MR图像的分类由ResNet18卷积神经网络(CNN)模型来实现。运用深度残差网络提升训练精度和测试精度。所使用的数据集包括19 146张7个前列腺MR序列图像(横断面T1WI、横断面T2WI、冠状面T2WI、矢状面T2WI、横断面DWI、横断面ADC、横断面PWI),选取其中2 800张图像作为训练集,选取剩余图像中的388张图像作为测试集。采用准确度评价ResNet18 CNN模型的效能。结果7个前列腺MR序列(横断面DWI、冠状面T2WI、横断面灌注成像、矢状面T2WI、横断面ADC、横断面T1WI和横断面T2WI)图像测试准确率分别为100.0%(44/44)、77.5%(31/40)、96.7%(116/120)、100.0%(44/44)、100.0%(44/44)、100.0%(52/52)和100.0%(44/44)。横断面PWI的分类0.8%(1/120)被错误地分到了横断面T2WI序列,仅2.5%(3/120)错误地分到矢状面T2WI序列;对于冠状面T2WI的分类15.0%(6/40)被错误地分到了横断面T2WI序列,7.5%(3/40)错误地分到矢状面T2WI序列。结论开发的能够自动分辨前列腺多序列MR图像的AI工具准确率高。

Objective To develop a convolution neural network (CNN) model to classify multi-sequence MR images of the prostate. Methods ResNet18 convolution neural network (CNN) model was developed to classify multi-sequence MR images of the prostate. A deep residual network was used to improve training accuracy and test accuracy. The dataset used in this experiment included 19 146 7-sequence prostate MR images (transverse T1WI, transverse T2WI, coronal T2WI, sagittal T2WI, transverse DWI, transverse ADC, transverse PWI), from which a total of 2 800 7-sequence MR images was selected as a training set. Three hundred and eighty eight 7-sequence MR images were selected as test sets. Accuracy was used to evaluate the effectiveness of ResNet18 CNN model. Results The classification accuracy of the model for transverse DWI, sagittal T2WI, transverse ADC, transverse T1WI, and transverse T2WI was as high as 100.0%(44/44,52/52), and the accuracy for transverse PWI was also as high as 96.7%(116/120). The accuracy for coronal T2WI was 77.5%(31/40). 0.8%(1/120) of transverse PWI was incorrectly assigned to transverse T2WI, and 2.5%(3/120) incorrectly assigned to sagittal T2WI. 15.0%(6/40) of coronal T2WI was incorrectly assigned to transverse T2WI, and 7.5%(3/40) to sagittal T2WI. Conclusion The experimental results show the effectiveness of our deep learning method regarding accuracy in the prostate multi-sequence MR images detection.