Arooshi Kumar, MD
@ArooshiK

Bathla G, Liu Y, Zhang H, Sonka M, Derdeyn C. Computed Tomography Perfusion–Based Prediction of Core Infarct and Tissue at Risk: Can Artificial Intelligence Help Reduce Radiation Exposure? Stroke. 2021;52:e755–e759.

It is well accepted that advanced CT perfusion (CTP) technology can help select patients who could benefit from endovascular treatment (EVT).1 However, CTP imaging is arterial input dependent, requires higher radiation than CT angiography, and utilizes timely and costly post-processing software.2,3 Authors Bathla et al. in this study explored if a machine learning algorithm, a convolutional neural network, which establishes complex relationships between many layers of visual imagery, could estimate cerebral blood flow and area-at-risk similar to the standard arterial dependent prediction method (RAPID).

Retrospective CTP data from 57 patients was split into training/validation (60%/40%) sets. The authors devised and validated separate U-net models, allowing for imaging segmentation, to predict core infarct (CBF) and tissue at risk (Tmax). Once trained, the full sets of 28 input images were sequentially reduced to equitemporal 14, 10, and 7 time points (tp) to further investigate if suboptimal arterial capture could be overcome by this predictive algorithm. The averaged structural similarity index measure (SSIM), a measure of similarity between two images, between the model-derived images and true perfusion maps was compared. For reference, the higher the SSIM, the better the reconstruction technique.

For the CBF model when compared to true images, the SSIM was 0.81±0.1 for the validation cohort when using all 28-tp. This decreased slightly to 0.80±0.1 (P=0.25) when CBF maps were generated using only 14-tp. For the Tmax model, the SSIM was 0.82±0.1 when using all 28-tp and 0.81±0.1 when using 14-tp (P=0.107). Of note, there was a modest correlation between model-derived and true CBF using all 28-tp at 0.69 and Tmax at 0.74. Figure 1 compared the RAPID prediction of CBF for a left middle cerebral artery territory infarct on top versus the Convolution Neural Network (CNN) based CBF prediction on the bottom.

Figure 1. Commercial software output versus CNN model output.
Figure 1. Commercial software output versus CNN model output.

This is an interesting proof-of-concept employment of CNN to predict CBF and Tmax, especially with suboptimal CTP data. The SSIM for both CBF and Tmax were similar whether using all the data with 28-tp or half the data with 14-tp. This suggests that despite incomplete data, the CNN reconstruction technique is decent. Furthermore, the authors realized that contrast arrival in intracranial vessels arrived around time point 7, suggesting much of the early time point data was likely not relevant to the model. Clinically, we are faced with challenges when the contrast bolus is suboptimal, and such augmented technology could help prevent repeat imaging and can lead to more timely interventions. The biggest caveat of this study is the low number of patients in the training set making its generalizability questionable. In the future, a larger study should be conducted to validate this proof-of-concept.                    

References:

  1. Albers GW, Marks MP, Kemp S, Christensen S, Tsai JP, Ortega-Gutierrez S, McTaggart RA, Torbey MT, Kim-Tenser M, Leslie-Mazwi T, et al; DEFUSE 3 Investigators. Thrombectomy for stroke at 6 to 16 hours with selection by perfusion imaging. N Engl J Med. 2018;378:708–718. doi: 10.1056/NEJMoa1713973
  2. Bathla G, Limaye K, Policeni B, Klotz E, Juergens M, Derdeyn C. Achieving comparable perfusion results across vendors. The next step in standardizing stroke care: a technical report. J Neurointerv Surg. 2019;11:1257–1260. doi: 10.1136/neurintsurg-2019-014810
  3. Zensen S, Guberina N, Opitz M, Köhrmann M, Deuschl C, Forsting M, Wetter A, Bos D. Radiation exposure of computed tomography imaging for the assessment of acute stroke. Neuroradiology. 2021;63:511–518. doi: 10.1007/s00234-020-02548-z