0983
Automatic AHA Classification of Carotid Atherosclerotic Lesions in Multicontrast MR Images using Deep Learning1Center for Biomedical Imaging Research, Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, People's Republic of China, Beijing, China, 2Vascular Imaging Laboratory, Department of Radiology, University of Washington, Seattle, WA, United States, Seattle, WA, United States
Synopsis
In this study, we aimed to develop a convolutional neural network (CNN) to classify carotid atherosclerotic lesions in high-resolution multicontrast MR images automatically using the modified American Heart Association (AHA) classification scheme as criteria. The network was trained on a large number of plaque images combined with lesion type labeled by experienced radiologists. Transfer learning was utilized to take the advantage of state-of-the-art CNN pre-trained on ImageNet dataset. The accuracy of lesion type classification achieved 85.1% with preprocessing and fine-tuning of the network.
Introduction
Multicontrast MRI can characterize the various components of carotid atherosclerotic plaque, which is crucial for lesion-type classification[1]. The modified American Heart Association (AHA) classification scheme of atherosclerotic plaque for MRI was concise and explicit, and extensively used for carotid plaque characterization[2]. However, it is time-consuming and subjective for radiologists to classify carotid atherosclerotic lesion type in multicontrast MR images. In this work, we explored the advanced CNN combining with transfer learning[3, 4] to classify carotid atherosclerotic lesions in multicontrast MR images automatically.Methods
Our dataset consists of 1,487 carotids (10-16 locations per carotid, and totally 19,720 locations) from 798 patients. Each location has four contrast weightings[5] (T1, T2, TOF and MPRAGE) with 256*256 in image size. The images of image quality ≤2 have been excluded. All data were labeled by experienced radiologists with five lesion types using CASCADE[6] according to modified AHA classification scheme (Figure 1). 80% of the data were randomly selected as training data, 10% as validation data and 10% as test data. (Figure 2)
The input of the network is four contrast MR images with the segmented carotid vessel wall, and the output is carotid atherosclerotic lesion type between 1-5 (Figure 3). We fine-tuned the residual network[7] with 101 layers pre-trained on ImageNet dataset[4] as our based model. We made four copy of the first convolution layer in the original residual network and connected each MRI contrast with one copy. The results of these four convolution layers were merged by summation, and connected to the rest of the net. We reduced the number of output to five classes. In addition, we flipped, rotated and cropped the image randomly as data augmentation in the training procession so as to improve its generalization capability. The modified model was trained and validated on the training and validation data set and tested on the test data set.
Results
Figure 4 shows the classification results of CNN compared with radiologists on the test dataset, and the proposed modal achieved accuracy of 85.1% (1679/1972). Figure 5 shows the sensitivity and specificity for each carotid atherosclerotic lesion type. The sensitivity and specificity, respectively, of CNN classification were as follows: type I lesion, 0.972 and 0.927; type II lesion, 0.660 and 0.944; type III lesion, 0.744 and 0.939; type IV lesion, 0.674 and 0.994; type V lesion, 0.575 and 0.985.Discussion
In this work, we proposed an automatic algorithm for carotid atherosclerotic lesion AHA type classification based on convolutional neural network (CNN) using multicontrast MR images. Through copying first convolution layer four times and summing them together, the CNN-based model could be utilized on the multi-contrast MR images. High sensitivity and specificity were achieved from type I lesion to type IV but poor performance for type V lesion.Conclusion
In this work, an automatic CNN-based classification of carotid atherosclerotic lesion type in multicontrast MR images has been implemented. Training and evaluation on a large-scale clinical data set shows that CNN-based model is capable of classifying intermediate to advanced atherosclerotic lesions in multicontrast MR images according to a modified AHA classification scheme.Acknowledgements
No acknowledgement found.References
1. Toussaint JF, LaMuraglia GM, Southern JF, et al. Magnetic resonance images lipid, fibrous, calcified, hemorrhagic, and thrombotic components of human atherosclerosis in vivo. Circulation. 1996;94:932–938.
2. Cai, Jian-Ming, et al. "Classification of human carotid atherosclerotic lesions with in vivo multicontrast magnetic resonance imaging." Circulation 106.11 (2002): 1368-1373.
3. Krizhevsky A, Sutskever I, Hinton G E. ImageNet Classification with Deep Convolutional Neural Networks[J]. Advances in Neural Information Processing Systems, 2012, 25(2):2012.
4. Russakovsky O, Deng J, Su H, et al. ImageNet Large Scale Visual Recognition Challenge[J]. International Journal of Computer Vision, 2015, 115(3):211-252.
5. Yuan, C., et al., In Vivo Accuracy of Multispectral Magnetic Resonance Imaging for Identifying Lipid-Rich Necrotic Cores and Intraplaque Hemorrhage in Advanced Human Carotid Plaques. Circulation, 2001. 104(17): p. 2051-2056.
6. D Xu, WS Kerwin, T Saam, M Ferguson, and C Yuan. Cascade: Computer aided system for cardiovascular disease evaluation. In Proc ISMRM, page 1922, 2004.
7. He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[J]. Computer Science, 2015.
Figures
