2831

Deep Learning-based Detection of DSC-Defined Penumbral Tissue on pCASL in Acute Ischemic Stroke
Kai Wang1, Qinyang Shou2,3, Samantha J. Ma1, David Liebeskind4, Xin J. Qiao5, Fabien Scalzo4, Jeffrey Saver4, Noriko Salamon5, and Danny JJ Wang1,4

1Lab of Functional MRI Technology, University of Southern California, Los Angeles, CA, United States, 2Stevens Institute of Neuroimaging and Informatics, University of Southern California, Los Angeles, CA, United States, 3Shanghai Jiaotong University, Shanghai, China, 4Neurology, UCLA, Los Angeles, CA, United States, 5Radiology, UCLA, Los Angeles, CA, United States

Synopsis

Arterial spin labeled (ASL) techniques can provide cerebral blood flow (CBF) measures without the use of a contrast agent, and it has been shown to provide largely consistent results with DSC perfusion in delineating hypoperfused brain regions in AIS while also providing information on hyperemic lesion. In this study, we develop a deep learning-based model to identify the hypoperfusion lesion on ASL images based on the DSC perfusion-defined penumbra region and diffusion weighted imaging (DWI). Our results show that deep learning can predict the DSC-defined penumbral region in ASL with dice coefficient=0.43.

Introduction

Acute ischemic stroke (AIS) patients with specific patterns of perfusion-diffusion lesion volume mismatch (i.e., penumbra) are more likely to benefit from endovascular thrombectomy1, 2. The main MR perfusion imaging method employed in AIS for determining the penumbra volume has been CT perfusion and dynamic susceptibility contrast-enhanced (DSC) MRI techniques, which requires X-ray exposure and/or injection of a contrast agent. Arterial spin labeled (ASL) techniques can provide cerebral blood flow (CBF) measures without the use of a contrast agent, and it has been shown to provide largely consistent results with DSC perfusion in delineating hypoperfused brain regions in AIS3. However, the precise delineation of hypoperfusion lesion and penumbra in ASL images remain challenging due to the low SNR and delayed arterial transit. In this study, we develop a deep learning-based model to identify the hypoperfusion lesion and penumbra in ASL images, using the perfusion and diffusion lesions in DSC perfusion weighted imaging (PWI) and diffusion weighted imaging (DWI) as supervision.

Methods

1. Data acquisition and processing

A total of 157 AIS patients underwent routine clinical stroke MRI on Siemens 1.5T Avanto or 3.0T TIM Trio systems using 12-channel head coils. Each patient had 1 to 4 MRI scans, resulting in a total of 174 usable image datasets. For CBF, a pseudo-continuous arterial spin labeling (pCASL) sequence with background suppressed 3D GRASE readout was applied with the following parameters: TR/TE/label time/post-labeling delay=4000/22/1500/2000ms; field of view=22cm; matrix size=64×64, 26×5mm slices; 30 pairs of label and control images with a scan time of 4min. Following motion correction, normalization (79*95*79, 2*2*2mm3) and pairwise subtraction between label and control images, quantitative CBF maps were calculated from the pCASL images. DSC images were acquired using a gradient-echo EPI sequence. Following motion correction and spatial smoothness, the Time-to-maximum (Tmax) map was generated using a cSVD deconvolution method by the commercial software OLEA, and a threshold>=6sec was used to identify the hypoperfusion regions. After thresholding, skull-stripping, CSF-masking, spatial smoothing and clustering were used for denoising and the resultant perfusion lesions (Tmax>=6sec) were used for training.

2. Network structure and training process

The network used in this project is highres3D4 ,which has 20 trainable layers with diluted convolution, diluting factor being 1, 2 and 4. Residual connections are employed for every two convolution layers. More details regarding the network can be found in Fig.1. The network was trained on 2 Nvidia GeForce GTX 1080 Ti GPUs. Volumes for training was randomly extracted from 3D preprocessed images, each volume is 48*48*48 and the batch size is 4. Augmentation on volume level were employed including rotation with a random angle in the range of [-10] for each of the three orthogonal planes and spatial rescaling with a random scaling factor in the range of [0.9,1,1]. The loss function is Cross Entropy, and Adam optimization method5 was used, learning rate lr=0.0001, $$$/beta $$$1 = 0.9 and $$$/beta $$$2 = 0.999. The total iteration number is 60,000, which allows the training process to reach the steady state. The training process was managed by NiftyNet6.

3. Result evaluation

Prediction output from the trained model was evaluated by comparing with ground truth. Dice coefficient was calculated for each subject, and precision/recall of each subject were also calculated. A precision-recall curve was plotted based on the output probability map, shown in Fig. 4. The ratio of training, validation and testing data was 8/1/1.

Results and Discussion

One representative case of prediction results was shown in Fig.2, along with the ground truth image, and different modalities of input images (ADC and CBF). The model performed well in capturing hypoperfused region in CBF images. However, the CBF hypoperfused region doesn’t match with region with prolonged Tmax in the DSC images perfectly, which may have caused the discrepancy of prediction and ground truth. Another case with perfusion lesion in the left posterior cerebral artery (PCA) territory was shown in Fig3. This is a case where our model didn’t give satisfying results, probably due to the lack of PCA lesions in the training data. The average dice coefficient is 0.43.

Conclusion

In general our model was able to find the hypoperfused region generated from Tmax images. Further research will be focused on improving the data cleaning procedure, and increasing the dataset so that more various input images can be used.

Acknowledgements

No acknowledgement found.

References

1. Albers GW, Thijs VN, Wechsler L, Kemp S, Schlaug G, Skalabrin E, et al. Magnetic resonance imaging profiles predict clinical response to early reperfusion: The diffusion and perfusion imaging evaluation for understanding stroke evolution (defuse) study. Annals of Neurology. 2006;60:508-517

2. Davis SM, Donnan GA, Parsons MW, Levi C, Butcher KS, Peeters A, et al. Effects of alteplase beyond 3 h after stroke in the echoplanar imaging thrombolytic evaluation trial (epithet): A placebo-controlled randomised trial. The Lancet Neurology. 2008;7:299-309

3. Wang DJ, Alger JR, Qiao JX, Hao Q, Hou S, Fiaz R, et al. The value of arterial spin-labeled perfusion imaging in acute ischemic stroke: Comparison with dynamic susceptibility contrast-enhanced mri. Stroke. 2012;43:1018-1024

4. Li W, Wang G, Fidon L, et al. On the compactness, efficiency, and representation of 3D convolutional networks: brain parcellation as a pretext task[C]//International Conference on Information Processing in Medical Imaging. Springer, Cham, 2017: 348-360.

5. Kingma, D., Ba, J.: Adam: A method for stochastic optimization (2014). arXiv:1412.6980

6. Gibson, Eli, et al. "NiftyNet: a deep-learning platform for medical imaging." Computer methods and programs in biomedicine 158 (2018): 113-122.

Figures

Fig.1 The structure of highres3D network. The network has three blocks, each contains six identical convolutional layers, with residual connections for every two layers. The first block uses conventional convolution, kernel size=3*3*3. This is to capture lower level features of the input volumes. The second and third uses diluted convolution, kernel size=3*3*3, diluted by 2 and 4 respectively. These deeper layers encode mid- and high-level image features. Image resolution is maintained throughout the network.

Fig.2 Prediction results of one subject compared with ground truth, along with input images. The prediction result matches well with ground truth. However, certain details are missing. The model seems to perform well in terms of capturing hypoperfused region shown in the CBF image, which is not exactly the same as prolonged Tmax region in DSC.

Fig. 3. Failed case for the model. The CBF image has a visually detectable hypoperfused region, which should be caused by PCA stroke, however the model completely missed it. On the one hand the hypoperfused region is relatively small, which is harder to catch; on the other hand, the model was given much less PCA stroke cases than MCA stroke cases, which increases the challenge.

Fig. 4. The average precison-recall curve of the inference set. The red dotted line is the baseline as our label ratio = 1:33.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
2831