A poor functional outcome of stroke patients may be predicted by early signs on CT and MRI. Radiology exams carried out within 6 hours after stroke onset can contain important insights supporting the prediction of post-thrombolysis hemorrhage of the infarct after thrombolysis.1 It is, therefore, of utmost importance to establish these signs as early and precisely as possible.
However, the early signs of a stroke on non-contrast CT (NCCT), and to a lesser extent MRI, are not straightforward to interpret using the naked eye. The most turned-to technique is NCCT, as it is the faster and more economical option. Hence, early stroke detection on NCCT could be an interesting area for artificial intelligence (AI) algorithms to offer their support and add value to the radiological part of the diagnosis. Although it is significantly easier to recognize early signs of brain ischemia on MRI, and MRI facilitates detection at an earlier stage of the disease process, AI could still play an important role by quantifying observations and removing inter-observer variance.
Below, we will discuss the most important early signs of a stroke on both NCCT and MRI. Additionally, we demonstrate the potential of AI by discussing a number of AI-based methods addressing the difficulties with an early image-based diagnosis of stroke and supporting image interpretation.
Early Signs of Stroke on NCCT
On NCCT, a stroke typically presents itself in the territory of the middle cerebral artery. For simplicity’s sake, we will discuss them in two groups: (semi) quantitative signs (e.g. as thirds of the middle cerebral artery) or qualitative signs (more of describing nature, e.g. obscuration of the lentiform nucleus).1–5
(Semi-) Quantitative Signs on NCCT
Ischemic cytotoxic edema can be observed on CT due to the failure of ion pumps, leading to edematous changes. A decreased attenuation of water in the brain will locally result in lower Hounsfield Units, characterizing the location of the infarct and allowing for (semi) quantitative measurements as described below.6
- Hypoattenuation in the basal ganglia
Hypoattenuation specifically in the basal ganglia indicates an early ischemic stroke. This sign can be observed within six hours after onset. In case of a middle cerebral artery (MCA) infarction, the basal ganglia are frequently involved.1,4,6
- Hypoattenuation of the vascular territory of the middle cerebral artery
Next to hypoattenuation of the basal ganglia, a decreased attenuation in the area of the MCA may be indicative of an infarct. A semi-quantitative scoring system is applied by determining whether less or more than one-third of the MCA flow territory is involved.1,2,4
- The ASPECT score
Technically speaking, the ASPECT (Alberta Stroke Program Early CT) score is not a sign, but a tool to analyze hypoattenuation in different brain regions. The ASPECT score is a number between 1 and 10 based on the number of affected brain regions. Each affected region gets appointed a score based on the level of hypoattenuation, giving an estimation of the functional outcome expected at 3 months. The lower the score, the worse the expected outcome is estimated to be.7 As the measurements of this score require manual assessment by the radiologist, it should be considered a semi-quantitative variable.1
Most early signs of a stroke are of a qualitative nature as there are no dedicated scoring systems involved. Below, we discuss the six most prominent qualitative signs of stroke on NCCT.
- Obscuration of the lentiform nucleus
The area of the lentiform nucleus may show decreased attenuation, leading to a less clear delineation between white and grey matter. This indicates cytotoxic edema and can be observed within 2 hours after stroke onset. Obscuration of the lentiform nucleus can also be referred to as “blurred basal ganglia”.2,6,8
- Loss of gray- and white-matter differentiation in the basal ganglia
This sign presents itself similarly to the “obscuration of the lentiform nucleus” and care should be taken not to confuse the two. It is characterized by decreased contrast and loss of precise delineation of the gray-white matter interface of all basal ganglia.2,9
- Cortical sulcal effacement
This expression of early ischemia on NCCT becomes apparent in the margins of the cortical sulci by expressing a decreased image contrast. The delineation between gray and white matter becomes less precise, caused by a localized mass effect.1,2,4
- General focal hypoattenuation
In addition to previously discussed hypoattenuated-related signs, discrete focal hypoattenuation can be detected by comparing brain structures to their contralateral counterparts. Additionally, assessing different locations within the same structure in search for abnormalities can support establishing increased local radiolucency.2,4
- The loss of insular ribbon, obscuration of Sylvian fissure sign
Fading discrete delineation of the gray-white matter interface at the lateral margin of the insula is often referred to as “the loss of the insular ribbon”.8,9 This manifestation is caused by hypodensity and swelling of the insular cortex.6
- The “dense artery sign”
Hyperattenuation of a brain vessel is called the “dense artery sign” and mostly occurs in the middle cerebral artery. This causes brightness on the CT images compared to any other (contralateral) artery or vein due to intravascular clot formation in the artery.2,3,9
Since quantitative signs can be extracted directly from the images, it is straightforward to use them to develop algorithms. Rule-based algorithms, which basically tell the computer step-wise what to do, are ideal to digitize the determination of quantitative scores related to stroke detection. An example is the aforementioned ASPECT score. Exploiting qualitative signs, on the other hand, requires the development of more complex methods. Qualitative parameters are great candidates for deep learning (DL) algorithms. By using large datasets as input, DL-related neural networks can learn to exploit parameters that are not necessarily used in the current diagnosis process. By deploying DL techniques, qualitative signs can be transformed into more objective quantitative measuring tools.
Early Signs of Stroke on MRI
MRI is a more time-consuming procedure; therefore, most standard stroke protocols rely on NCCT. However, since MR imaging can offer additional information compared to NCCT images9–12, a number of early signs of stroke as detected on MRI are discussed below.
- Signal intensity changes on diffusion-weighted imaging (DWI) MRI
Diffusion-weighted imaging (DWI) MRI is able to uncover changes in signal intensity in the stroke-affected area within minutes after stroke onset (for more information on this technique, read our blog on our blog on measuring TSS). Hence, the DWI sequence offers a valuable tool in addition to NCCT due to the short time frame in which it is able to detect ischemia. Furthermore, the sensitivity of DWI outperforms the sensitivity of most NCCT-based signs.13
- High-intensity zone on FLAIR images
After an ischemic event, fluid-attenuated inversion recovery (FLAIR) images show no changes in signal intensity in the first 3 hours. After that period, the ischemic infarction zone will develop a high signal intensity zone. FLAIR imaging, therefore, presents itself as a suitable method to determine the time of onset of the suspected ischemia.10,14 Additionally, T1- and T2-weighted images can be used to assess stroke cases. In the case of a large stroke flow, voids can be observed in the large arteries, as the T2W image will display areas of increased focal signal intensity. T1W images show signal intensity changes in case of ischemia. Unfortunately, T1W analysis and T2W analysis can only be performed 16 and 8 hours respectively after the onset of a stroke, dismissing their added value to the case of acute stroke detection.12
Using MRI to determine TSS
As treatment trajectory depends strongly on time since stroke onset (TSS), quantification of TSS is of the utmost importance. MRI enables TSS measurement by combining DWI and FLAIR images and by determining the mismatch between both. Learn more about this technique in our blog on measuring TSS.14–16
In 3 Steps Towards AI-Based Software for Early Stroke Detection
Artificial intelligence algorithms can assist stroke detection in various ways. Different solutions are discussed in the remainder of this article, entrusting to the computer a smaller or bigger role in the process.
- Highlighting what is already there
The most straightforward way in which AI could assist clinical practice is by reading images and making possible abnormalities stand out more, and therefore easier to assess for a radiologist. An example of such an algorithm, solely magnifying the presence of stroke on NCCT, is proposed by Przelaskowski et al. In their approach, which uses a wavelet-based image processing method, hypodense areas in the images are enhanced, making it easier for the radiologists to detect possible infarcts. Using their approach on CT images that were obtained between 1 and 5 hours after stroke onset, they found a substantial gain in sensitivity, which increased from 12.5% to 56.3%.17
- Solving pieces of the puzzle as we know it
For every relevant question you would like to answer in case of a stroke suspicion, you can try to develop an algorithm. In other words: you teach the computer to do what radiologists would do in the way radiologists are used to doing it. For example, for automated delineation, measurement of infarct volume, and thrombus detection and measurements. An algorithm would add value by decreasing the time investment required and by eliminating inter- and intra-observer variance.18–20
A range of examples of such algorithms can be found in digitizing familiar scores used for stroke assessment. For example, the ASPECT score for outcome prediction. Takahashi et al. developed a computer-aided detection scheme to identify which patients are eligible for treatment with tPA. The algorithm used linear discriminate analysis to classify all ASPECT regions as hypo-attenuated or normal based on NCCT images obtained within 6h after stroke onset. Results showed an average accuracy of 85% for the hypoattenuation classification per patient, with the assessment of an experienced neuroradiologist as a ground truth.21
- Solving the problem the AI way from scratch: machine learning
Lastly, one can deploy machine learning (ML) or DL to extract more valuable information from stroke images than radiologists are currently able to do with the naked eye. By presenting a large volume of manually labeled images to the computer, the algorithm will learn to associate the treatment outcome to certain image characteristics. Interestingly, if sufficient data is presented, ML algorithms can learn to solve this problem “in its own way”, not necessarily utilizing the same image characteristics as a radiologist.
Bentley et al. developed an ML algorithm for diagnosis support by predicting the development of spontaneous intracerebral hemorrhage (SICH) after administering tPA to stroke patients. They trained an algorithm using 116 acute ischemic stroke patients treated with intravenous thrombolysis, of which 16 developed SICH. The ML software was able to identify 9 out of 16 SICHS on CT brain images, compared with detecting 1-5 SICH cases by standard prognostic scores (e.g. SEDAN and HAT score).22 A promising start for developing a powerful tool to help prevent additional brain damage.
Currently, the detection and assessment of stroke are strongly qualitative in nature. However, more quantitative diagnosis processes could be of great value. They would improve the objective character of the assessment, next to offering the possibility of detecting smaller abnormalities. AI algorithms have great potential to contribute to such a shift. Additionally, the software can speed up the process or provide more differentiated information.
Are you curious about what Quantib is doing related to stroke diagnosis? Check out the webpage of our CASE project and learn more about the implementation of the stroke decision support system we are developing together with QMENTA.
- Wardlaw, J. M. & Mielke, O. Early Signs of Brain Infarction at CT: Observer Reliability and Outcome after Thrombolytic Treatment—Systematic Review. Radiology 235, 444–453 (2005).
- Provenzale, J. M., Jahan, R., Naidich, T. P. & Fox, A. J. Assessment of the Patient with Hyperacute Stroke: Imaging and Therapy. Radiology 229, 347–359 (2003).
- Pressman, B. D., Tourje, E. J. & Thompson, J. R. An early CT sign of ischemic infarction: Increased density in a cerebral artery. Am. J. Roentgenol. 149, 583–586 (1987).
- Kjos, B. O., Brant Zawadzki, M. & Young, R. G. Early CT findings of global central nervous system hypoperfusion. Am. J. Neuroradiol. 4, 1043–1048 (1983).
- Jordan, Y. J., Jordan, J. E., Lightfoote, J. B. & Ragland, K. D. Quality outcomes of reinterpretation of brain CT studies by subspecialty experts in stroke imaging. Am. J. Roentgenol. 199, 1365–1370 (2012).
- Thurnher, M. Radiology Assistant. Brain Ischemia - Imaging in Acute Stroke (2088). Available at: http://radiologyassistant.nl/en/p483910a4b6f14/brain-ischemia-imaging-in-acute-stroke.html.
- Gillard, F. & Sair, H. Radiopaedia. Alberta stroke program early CT score (ASPECTS) Available at: https://radiopaedia.org/articles/alberta-stroke-program-early-ct-score-aspects-1.
- Tomandl, B. F. et al. Comprehensive Imaging of Ischemic Stroke with Multisection CT. RadioGraphics 23, 565–592 (2003).
- Lum, C. ing of Acute Stroke 1 OBJECTIVES. Perfusion 48109, 75–96 (2006).
- Makkat, S. et al. Signs of acute stroke seen on fluid-attenuated inversion recovery MR imaging. Am. J. Roentgenol. 179, 237–243 (2002).
- 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).
- 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).
- Barker, P. B. et al. Acute stroke: evaluation with serial proton MR spectroscopic imaging. Radiology 192, 723–732 (1994).
- 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).
- Legrand, L. et al. Do FLAIR vascular hyperintensities beyond the DWI lesion represent the ischemic penumbra? Am. J. Neuroradiol. 36, 269–274 (2015).
- Mair, G. & Wardlaw, J. M. Imaging of acute stroke prior to treatment: current practice and evolving techniques. Br. J. Radiol. 87, 20140216 (2014).
- Przelaskowski, A. et al. Improved early stroke detection: Wavelet-based perception enhancement of computerized tomography exams. Comput. Biol. Med. 37, 524–533 (2007).
- Gillebert, C. R., Humphreys, G. W. & Mantini, D. Automated delineation of stroke lesions using brain CT images. NeuroImage Clin. 4, 540–548 (2014).
- Boers, A. M. et al. Automated cerebral infarct volume measurement in follow-up noncontrast CT scans of patients with acute ischemic stroke. AJNR. Am. J. Neuroradiol. 34, 1522–7 (2013).
- Löber, P. et al. Automatic thrombus detection in non-enhanced computed tomography images in patients with acute ischemic stroke. i, (2017).
- Takahashi, N. et al. Computer-aided detection scheme for identification of hypoattenuation of acute stroke in unenhanced CT. Radiol. Phys. Technol. 5, 98–104 (2012).
- Bentley, P. et al. Prediction of stroke thrombolysis outcome using CT brain machine learning. NeuroImage Clin. 4, 635–640 (2014).