Deep learning for the detection of moyamoya angiopathy using T2-weighted images: a multicenter study
Maoxue Wang1,2,3, Song Luo4, Chaoyong Xiao5
1Department of Radiology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
Background:
Moyamoya angiopathy (MMA) can be potentially missed in the initial magnetic resonance (MR) examination without MR angiography (MRA). The aim of this study was to develop an optimal deep learning model based on T2-weighted imaging (T2WI) for MMA detection.
Methods:
This retrospective multicenter study included MMA patients, control group patients with normal MRA and patients with cerebrovascular disease except MMA from seven hospitals (site 1 to site 7). Five models, namely shallow convolutional neural network (SCNN), LeNet-5 Convolutional Neural Network (LeNet), Visual Geometry Group Network (VGG), Residual Neural Network (ResNet) and Dense Convolutional Network (DenseNet), were used for training and validation. The model training and internal validation were performed with data from sites 1-4. Data from sites 5-7 were used for independent external validation, and the optimal model was selected according to the results of accuracy. Chi-squared test was used to further verify the influence of different MR manufacturers, field strength, age at the MR examination and MRA score on the optimal model.
Results:
A total of 1,038 MMA patients, 1,211 normal MRA and 271 patients with cerebrovascular disease except MMA were included. DenseNet showed the highest accuracy (0.859, 95% CI: 0.833, 0.884) in the independent external validation, which was not significantly different from that of VGG (0.834, 95% CI: 0.807, 0.861) and ResNet (0.855, 95% CI: 0.829, 0.880) but was significantly higher than that of SCNN (0.631, 95% CI: 0.595, 0.665; P<0.001) and LeNet (0.563, 95% CI: 0.527, 0.599; P=0.001). The accuracy of 1.5 T data was higher than 3.0 T data (χ2=6.559, P=0.01). The accuracy of MMA with MRA who scored more than 5 was higher than that scoring ≤5 (≤5 vs. 6-10: χ2=10.734, P=0.001; ≤5 vs. ≥11: χ2=10.369, P=0.001).
Conclusions:
DenseNet based on T2WI can be used to screen MMA, outperforming SCNN, LeNet, VGG and ResNet. The MRA score of MMA affected the DenseNet accuracy.
