Yajing Zhang1, Weiwei Jiang1, Zhizheng Zhuo2, Yunyun Duan2, and Yaou Liu2
1Philips Healthcare, Suzhou, China, 2Beijing Tiantan Hospital, Capital Medical University, Beijing, China
Synopsis
Multiple
sclerosis (MS) and neuromyelitis optical spectrum disorders (NMOSD) are common demyelination
diseases in central neural system. With increasingly used MR examinations in
clinical practice, the detection and delineation of white matter lesions is
helpful to the calculation of total lesion volume, which is of great
significance to clinical treatment planning. However, manual delineation is
time-consuming to the clinicians and may lead to poor repeatability. In this
work, we employed deep learning method to establish a tool for WM lesion
delineation of MS and NMOSD routine MRI data from multiple centers.
Introduction
MRI has been
widely used to detect the white matter lesions and it is important to
accurately delineate the lesion and measure the lesion volumes for treatment
planning. However, manual delineation by radiologists or clinicians is
time-consuming and the intra- and inter-rater reproducibility is not guaranteed
due to the intensive workload and the different perspectives from raters. This
study aims to establish a deep learning (DL) based automated white matter
lesion detection and delineation tool to assist the WM lesion identification
work for radiologists, which could be applied to MR images from different
clinical centers with MS or NMOSD lesions and with wide range of venders, image
resolutions and slice thickness.Methods
The MS and NMOSD data
were collected from 6 clinical centers (Table 1). The clinical routine T2 FLAIR
were acquired on 3T scanners from multiple vendors (Philips, GE and Siemens). Each
patient data contained 17-30 slices to cover the whole brain area. A total of 237
MS patients and 137 NMOSD patients were included in this multi-center study. Manual
labeling of the white matter lesions were performed by experienced radiologists
with cross-check. Original data with in-plane resolution ranging from 0.4296×0.4296mm2/px
to 0.75×0.75 mm2/px, and slice thickness ranging from 3mm to 15mm were
resampled to matrix of 256×256 before network training. 80% of the patient data
were arranged as training data from each clinical center in a balanced way. The
rest were used for testing.
We employed four
DL network strategies for the neural networks: FCN (fully convoluted network 1)
with upsample stride 16, 8 and 4 predictions back to pixels in a single step; and
the Unet 2 strategy. The transferred learning was used, with
implemented initial values from the Image-Net. For all the four networks, the
encoder was used as VGG-19 3, and the loss function as the mean
cross entropy, with mini-batch based stochastic gradient descent 4
and Adam optimizer 5. The learning rate was set to 3e-5. The
network was implemented through Tensorflow, with GPU 1.4GHz, 1080Ti.
To quantitatively
evaluate the performance of the networks, we calculated the prediction
accuracy, which was defined as Accuracy=(TP+TN)/N, where the TP
(true positive) was the lesion detection cases with Dice ratio larger than 0.5;
TN (true negative) was the healthy slices with no lesion detected. The accuracy
was calculated for the prediction from each clinical center. Results
Figure 1 demonstrated the lesion
detection from 4 DL network strategies. For the consecutive slices with MS lesions
manually delineated, all the 4 networks could segment the white matter lesions consistently
with the ground truth. The WM lesion for NMOSD case was shown in Figure 2. The
small sized WM lesion in NMOSD consecutive slices were delineated consistently
with the ground truth. Table 2 quantified the lesion detection accuracy of the
4 strategies for each disease type and from each data source center. The Unet
achieved overall accuracy of 87.66% for MS and 89.77% for NMOSD, superior to
the other 3 strategies, and achieved superior accuracies in the dataset from
most centers. The accuracy on the whole testing set reached 88.64% with Unet
strategy. The processing time for WM lesion delineation
on one patient (with 30 slices) was within 10s.Discussion and Conclusion
We have established an automated white matter lesion detection and
delineation tool using DL network. Through the comparison of 4 DL network
strategies, the Unet outperformed with regard to the detection accuracy and it can
be applied to a wide spectrum of MRI data in both MS and NMOSD data. Future
work includes refining the DL network and improving the practical workflow.Acknowledgements
No acknowledgement found.References
1. Long J, Shelhamer E,
Darrell T. Fully convolutional networks for semantic segmentation. Proc IEEE Conf
CVPR. 2015: 3431-3440.
2. Ronneberger O, Fischer P,
Brox T. U-net: Convolutional networks for biomedical image segmentation. Int.
Conf. Med. image computing and computer-assisted intervention. Springer, Cham,
2015: 234-241.
3. Simonyan K, Zisserman A.
Very deep convolutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556, 2014.
4. H. Robinds and S. Monro. A
stochastic approximation method. Annals of Mathematical Statistics, 1951; 22:
400–407.
5. Adam Kingma, D. P.,
Ba, J. L.. Adam. A Method for Stochastic Optimization. Int. Conf. Learning
Representations, 2015: 1–13.