Introduction

The proportion of red blood cell (RBC) content in stroke thrombi is associated with valuable clinical parameters including stroke etiology and responsiveness to thrombectomy therapy.^1,2 Thrombus RBC content is currently only determined retrospectively through histological examination. Previous work has shown that R₂^* and quantitative susceptibility mapping (QSM) values are sensitive to blood clot RBC content in vitro;^3,4 this study examines the feasibility of applying deep learning to these quantitative MR images for predicting RBC content in human stroke thrombi ex vivo.

Methods

Thrombus collection- Acute ischemic stroke patients treated with EVT at the local stroke centre between February 2016 to November 2017 were recruited for this study. 48 acute stroke thrombi were scanned ex vivo within 6 hours of retrieval and fixed in formalin for subsequent histological examination.
Imaging- Thrombi were imaged at 3T with a 32-channel head-coil using a custom dual echo-train 3D GRE sequence⁵ (TE₁/∆TE/TE₅ = 3.20/1.46/9.04 ms, TE₆/∆TE'/TE₁₀ = 16.75/7.15/45.35 ms, TR: 47.6 ms, resolution: 0.94 x 0.94 x 1 mm³, matrix: 192 x 192 x 36, BW: 142.86 kHz, flip angle: 10°, scan time = 5 min 33 sec). A naturally co-registered balanced steady-state gradient echo sequence (FIESTA-C) with identical resolution was also acquired to facilitate thrombus segmentation (TE = 3 ms, TR: 6.1 ms, flip angle: 40°, phase cycles = 4, scan time = 2 min 47 sec).
Image post-processing- Raw image data were reconstructed using the Orchestra Software Development Kit (GE Healthcare; Milwaukee, WI) in Matlab. Channel-combined complex GRE data were processed using the B0-NICE⁶ and MEDI QSM⁷ algorithms to calculate R₂^* and QS maps, respectively. Late echo (TE = 31 ms) magnitude images were generated through sum of squares coil combination. Thrombus segmentation was performed semi-automatically on the FIESTA-C images using house-built Matlab code.
Histology- Fixed thrombi were sectioned at 5 μm slice thickness and stained with H&E. The ImageJ color segmentation plugin (EPFL, Lausanne, Switzerland) was used to quantify histological thrombus RBC content.
Deep learning network- A 3-layer convolutional neural network (CNN) implemented in Matlab was trained to predict histological RBC content from segmented, normalized 3-channel (R₂^*, QSM, late-echo GRE) thrombus MR image slices. A schematic diagram of the CNN is shown in Figure 1. 188 thrombus MR slices derived from the 48 thrombi were available for analysis; it was assumed that thrombus RBC content was consistent across all MR slices. Data augmentation involving equalization of input slices based on RBC content (true/false), random geometric transformation (true/false), and dataset duplication (×1 – ×5; prior to geometric augmentation) were iteratively applied. A flowchart depicting the data augmentation scheme is shown in Figure 2. At each data augmentation setting, the network was evaluated across a range of batch sizes, training epochs and dropout factors. Network parameters and data augmentation specifications are detailed in Table 1. 8-fold cross validation was used to evaluate each network’s performance, average accuracy (predictions within 10% of histology), absolute error and root mean square error (RMSE) across all folds were recorded. All slices associated with any particular thrombus were used for either network training or testing only within any fold.

Results

Mean thrombus RBC content determined by histological analysis was 39 ± 13%. Average accuracy and absolute error of the top-5-performing networks at each data augmentation setting are shown in Figure 3. Average network accuracy and absolute error improved from 65±1% and 8.6±0.2% when trained on the original, unaugmented dataset up to 71±1% and 8.1±0.2%, respectively, after augmentation. Root mean square error improved only marginally with augmentation, decreasing from 10.2±0.3% in the original dataset down to 10.0±0.2% after augmentation.

Discussion

Despite a small dataset available for training, the CNN was able to accurately predict thrombus RBC content. Network performance improved as more drastic data augmentation was applied, in particular performance improved only after input equalization, geometric transformation and dataset duplication were all applied. Increasing dataset duplication improved network performance up to a factor of ×3, after which no further improvement was seen; dataset duplication may improve network performance only up to a point, but implies the network would show improved performance if trained on a larger thrombus dataset. Nonetheless, the mean absolute error in RBC content prediction of approximately 8% is already smaller than the average difference in RBC content between stroke etiologies (~14%)^1,8 or between thrombi responsive vs. non-responsive to EVT (~17%),^9,10 suggesting that a network analyzing stroke thrombi in vivo with similar performance would provide sufficient accuracy to guide acute stroke treatment.

Conclusion

The CNN learned to accurately predict stroke thrombus RBC content from quantitative, multiparametric MR data, and performance improved through dataset augmentation. These results suggest this method may be useful for predicting thrombus RBC content and informing acute stroke treatment decisions in stroke patients in vivo.

Acknowledgements

The authors wish to thank Justina Diaz Legaspe for recruiting and consenting participants, Ada Manini, Leanne Sperlich, Jon Collier, Steve Eggett, Sarah Sitts and Barb Lehrbass for storage of thrombi and facilitating sample transfer, the interventional neuroradiologists at London Health Sciences Centre for thrombus retrieval, Caroline O'Neil for performing thrombus histology, and Trevor Wade for implementation of the pulse sequence.