How can AI help determine time of ischemic stroke onset?

The prescribed moment to perform thrombolysis is within 3 to 4.5 hours after the onset of stroke (the guideline of the Dutch Neurology, Radiology and IR Association), since administration outside of this time window leads to an increased risk of intracranial bleeding.1 Therefore, at the acute setting of stroke, one major contraindication to thrombolysis is an unknown stroke onset time. Depending on, for example, the stroke location, a less strict interpretation of the cutoff time can be applied. Anterior brain circulation requires a cutoff of 4.5 hours, and even 6 hours is occasionally applied if the thrombolysis is administered intra-arterially during interventional procedures.2

During sleep, stroke symptoms are not noticed by the patient. However, up to 25% of stroke patients are reported to have been sleeping during the onset of a stroke. As this complicates the exact determination of time since stroke (TSS), these patients have been excluded from thrombolytic trials.3 However, Elliot reported that most events of stroke occur between 6 AM and noon, suggesting that people waking up in the morning with stroke symptoms are likely to have a TSS of less than a few hours.4,5 This only strengthens the burning platform for determination of TSS in (preparing for) treatment of ischemic stroke. How are we to determine whether the patient’s TSS is 3 to maximum 4.5 hours?

Five MRI sequences to help you determine TSS

The answer lies in multiparameter MRI. Diffusion-weighted imaging (DWI) MRI was developed in the mid-1990s, but it took some years to become a widespread method.6 DWI MRI is capable of detecting signal intensity changes in the stroke-affected area of the brain within minutes of the onset, hence this imaging sequence can be very valuable in the determination of TSS.7

Second sequence candidate is spectroscopy. Spectroscopy analyses the chemical spectrum of the investigated brain cells. Twenty-four hours after TSS, a significant accumulation of lactate can be observed in the affected ischemic brain cells. Doing the same measurement only 3 hours after TSS, a very subtle difference can be seen. However, this is not sufficiently convincing to decide for or against thrombolysis.8

Thirdly, T2-weighted sequences can be deployed. In case of a large stroke, flow voids can be seen in the large arteries. Unfortunately, the changes in signal intensity (increased signal) can only be observed 8 hours after the start of the ischemic changes, hence T2 cannot safely be used to determine the clinical thrombolysis window as measurement accuracy exceeds 6 hours.2

As a fourth option, T1-weighted sequences could be used, however, as T1 signal intensity changes (decreased signal) only 16 hours after TSS, they prove even less useful than T2-weighted sequences.2

However, the most likely candidate for TSS determination is the combination of fluid-attenuated inversion recovery sequence (FLAIR) and the DWI sequence. After an ischemic event, FLAIR images show no changes in signal intensity in the first 3 hours.9 Therefore, an anatomical mismatch between the increased signal intensity in DWI-MRI images and the area with physiological signal intensity of the FLAIR sequence marks the infarct zone with an onset of 3 hours or less and thus discerns the intravenously thrombolysis candidates and non-candidates in a very safe margin. Research has shown quite varying statistics on the results. Specificity has been shown to be in the range of 64-93%, sensitivity between 48-92%.9,10 However, if the mismatch is observable, it is an interesting case for artificial intelligence algorithms.

What about CT?

It has been shown that during the first 12 hours after stroke onset DWI MRI shows a higher sensitivity than CT, during the first 6 hours the sensitivity of CT is merely 40%.11 More recent research tested the ability of CT to quantify water uptake, supporting the estimation of TSS. Authors claim a sensitivity and specificity of over 90% for differentiation between TSS <4,5h and TSS>4,5h, making this an interesting method for further investigation.12

How can AI help?

Identifying the time of the onset of stroke using artificial intelligence is a relatively unexplored area. Stroke diagnosis related research has mainly focused on automatic lesion segmentation or determination of the severity of the lesion and its location, e.g. by implementing the ASPECT score.13–17 For more information on our stroke diagnosis support system, see our CASE project webpage.

Work by Lee et al. shows a first attempt at estimating TSS using a machine learning approach. Based on the DWI-FLAIR mismatch, they developed an automated system allowing for classification of acute ischemic stroke patients into three groups: within 4.5 hours, between 4.5 and 6 hours and more than 6 hours of the onset of symptoms. Unfortunately, no comparison to radiologist performance is included in this study, making it hard to estimate the additional value of a machine learning algorithm. Also, a small dataset of only 149 patients was used, raising questions about the wider applicability of this algorithm.18

Ho et al. applied a deep learning approach to perfusion-weighted images (PWI), DWIs, FLAIR images and ADC maps. Their aim was to determine whether TSS was less than 4.5 hours or more than 4.5 hours. Algorithm performance showed an improvement in number of true positives and reduction in false positives versus results from previous clinical research using DWI-FLAIR mismatch for TSS determination.19  

So now what?

As thrombolysis treatment is preferably administered within 3, maximum 4.5 hours of stroke onset, the above-mentioned research provides an interesting starting point for accuracy improvement of TSS determination. However, further refinement of the algorithms and further validation is necessary to develop a robust TSS determination algorithm.

Read our blog on The role of CT and MR in stroke patients to learn more about stroke imaging. 

Bibliography

  1. Schonewille, W. J., Dijk, E. J. Van, Boiten, J. & Dippel, D. W. Behandeling van het acute herseninfarct via de veneuze en arteriële route. Ned. Tijdschr. Geneeskd. 154, 1–7 (2010).
  2. Allen, L. M., Hasso, A. N., Handwerker, J. & Farid, H. Sequence-specific MR Imaging Findings That Are Useful in Dating Ischemic Stroke. RadioGraphics 32, 1285–1297 (2012).
  3. Fink, J. N. et al. The Stroke Patient Who Woke Up. Stroke 988–993 (2002).
  4. Elliott, W. J. Circadian Variation in the Timing of Stroke Onset : A Meta-analysis. 992–996 (1998).
  5. Petkova, M. et al. MR Imaging Helps Predict Time from Symptom Onset in Patients with Acute Stroke: Implications for Patients with Unknown Onset Time. Radiology 257, 782–792 (2010).
  6. Schaefer, P. W., Grant, P. E. & Gonzalez, R. G. Diffusion-weighted MR Imaging of the Brain. Radiology 217, 331–345 (2000).
  7. Beauchamp, N. J., Ulug, A. M., Passe, T. J. & van Zijl, P. C. MR diffusion imaging in stroke: review and controversies. RadioGraphics 18, 1269–1283 (1998).
  8. Barker, P. B. et al. Acute stroke: evaluation with serial proton MR spectroscopic imaging. Radiology 192, 723–732 (1994).
  9. Thomalla, G. et al. Negative fluid-attenuated inversion recovery imaging identifies acute ischemic stroke at 3 hours or less. Ann. Neurol. 65, 724–732 (2009).
  10. Legrand, L. et al. Do FLAIR vascular hyperintensities beyond the DWI lesion represent the ischemic penumbra? Am. J. Neuroradiol. 36, 269–274 (2015).
  11. Mullins, M. E. et al. CT and conventional and diffusion-weighted MR imaging in acute stroke: study in 691 patients at presentation to the emergency department. Radiology 224, 353–360 (2002).
  12. Minnerup, J. et al. Computed tomography-based quantification of lesion water uptake identifies patients within 4.5 hours of stroke onset: A multicenter observational study. Ann. Neurol. 80, 924–934 (2016).
  13. Maier, O., Schröder, C., Forkert, N. D., Martinetz, T. & Handels, H. Classifiers for ischemic stroke lesion segmentation: A comparison study. PLoS One 10, 1–16 (2015).
  14. Chen, L., Bentley, P. & Rueckert, D. Fully automatic acute ischemic lesion segmentation in DWI using convolutional neural networks. NeuroImage Clin. 15, 633–643 (2017).
  15. Pustina, D., Coslett, H. B., Turkeltaub, P. E., Tustison, N. & Schwartz, M. F. Automated segmentation of chronic stroke lesions using LINDA: Lesion Identification with Neighborhood Data Analysis. Hum. Brain Mapp. 37, 1405–1421 (2016).
  16. Herweh, C. et al. Performance of e-ASPECTS software in comparison to that of stroke physicians on assessing CT scans of acute ischemic stroke patients. Int. J. Stroke 11, 438–445 (2016).
  17. Nagel, S. et al. e-ASPECTS software is non-inferior to neuroradiologists in applying the ASPECT score to computed tomography scans of acute ischemic stroke patients. Int. J. Stroke 12, 615–622 (2017).
  18. Lee, H., Ham, S., Lee, E., Kim, N. & Kang, D. A machine learning approach to identify acute stroke patients within 4.5h or 6h from symptom onset. (2017). Available at https://files.aievolution.com/hbm1701/abstracts/36867/3246_Lee.pdf.
  19. Ho, K. C., Speier, W., El-Saden, S. M. & Arnold, C. W. Classifying Acute Ischemic Stroke Onset Time using Deep Imaging Features. Available at https://amia2017.zerista.com/event/member/389179.