Feature Article

Deep Learning and Stroke Imaging: Predicting and Improving Outcomes

Time is of the essence in stroke detection and management. Identifying and combining acute imaging features is essential to accurately predict the final lesion volume in patients sustaining acute ischemic stroke. Indeed, several randomized clinical trials showed superior clinical outcomes based on neuroimaging criteria.1

However, the correlation between neuroimaging data and prediction of treatment results holds the key to successful outcomes.

The Role of Deep Learning and Neural Networks in Stroke Management

Application of deep learning (i.e., artificial intelligence) techniques employing deep neural networks (DNNs) to stroke imaging data, has opened up new vistas into acute intervention and prognosis of stroke.

Although stroke is generally detected initially based on clinical symptoms, computer-assisted diagnosis using deep learning approaches serves as an adjunct for rapid diagnosis. It ensures that the initial CT imaging captures the event with unerring accuracy. Deep learning techniques increase the diagnostic and treatment efficacy and accuracy. They represent a promising and personalized therapeutic tool for patients diagnosed with ischemic stroke. Indeed, artificial intelligence can be used to spontaneously generate natural language captions describing images for rapid diagnosis and treatment of stroke.1

Neural networks are mathematical and computational models designed to mimic human brain. Comprising interconnected nodes that facilitate information processing using a computational approach, they are used to analyze and compare large volumes of data samples to detect correlations. Indeed, complex neurological and mental phenomena can be modeled using an interconnected network of simple uniform units. The structure of neural networks is altered depending on external and internal data acquired during the "training" phase. Neural networks are currently used to perform various tasks based on a complex relationship between data inputs and outputs or patterns.

Deep neural networks (DNNs) facilitate segmentation of images, as well as radiomic applications in which raw signals are modeled and automated, and prognostication using multiple modalities. Convolutional neural networks (CNNs), which are the group of DNNs involving images, are useful in acute stroke management and prognosis. CNNs aid in stroke imaging by converting raw images into features that can be utilized to train subsequent models. In contrast to the manual delineation of regions of interest in radiology images practiced currently, deep learning approaches, including CNNs allow an automated prediction of tissue fate. For example, deep learning strategies have been used to predict altered recovery following stroke MRI fluid-attenuated inversion based on acute diffusion-weighted imaging maps.2

A combination of deep learning-based models with image-based CNNs may be highly accurate in terms of prognosis compared with current models, resulting in potential clinical benefits as well as aiding in patient selection during clinical trial interventions.1

Prediction of Therapeutic Outcomes in Acute Ischemic Stroke

Anne Nielson and colleagues at Aarhus University Department of Clinical Medicine, Center of Functionally Integrative Neuroscience and MINDLAB, in Denmark have developed and validated a model for prediction of tissue outcome and evaluation of therapeutic effect in acute ischemic stroke using deep learning techniques employing neural networks.3

The researchers used acute magnetic resonance imaging to develop and train a deep convolutional neural network (CNNdeep) to forecast final imaging outcome. They treated 222 patients including 187 exposed to recombinant tissue-type plasminogen activator (rtPA). They compared the performance of deep and shallow CNNs based on several different imaging biomarkers. The control group of patients was not exposed to thrombolytic therapy and was evaluated to determine the treatment impact.3

The results indicated a significantly better performance of the deep CNN in predicting the final treatment outcome compared with the other models tested, and was markedly but not significantly better than the shallow neural network.3

“The use of several biomarkers/features from magnetic resonance imaging is rather innovative and is an approach that reduces bias and variability intrinsic in deep learning methods – that of ground-truth disagreement,” commented John Kalafut, Director Architecture and Workflow at GE Healthcare Digital Solutions. “One can argue there is still the potential for bias because of human annotation via markup and delineation of infarct area on the T2-FLAIR data but the interesting results here show the sensitivity of the various biomarkers,” Kalafut added.

The Danish researchers found that the deep CNN significantly enhanced the predictions and was superior to all the other modalities investigated. It also facilitated the differentiation of outcomes based on treatment strategy, with significant differences in the volume of final infarct. “The considerable prediction improvement accuracy over current state of the art increases the potential for automated decision support in providing recommendations for personalized treatment plans,” the Danish researchers concluded.3

However, the data may be very preliminary, and a larger cohort study is needed to validate the findings, according to Dr Kalafut. The approach needs to be compared with deep learning-based methods that directly segment the infarct itself using apparent diffusion coefficient.

The multiple biomarkers used by Nielson and colleagues may be appropriate for various radiomic applications in which textural and feature analysis of images contributes to varying degrees of lesion classification or prediction. In general, the Danish study continues the development and knowledge base highlighting the role of physiologic imaging in the assessment, identification and treatment of acute stroke. It also demonstrates the utility of MRI in patient management.

Use of Image-based Algorithms in Ischemic Stroke

Previous studies accurately predicted the 90-day motor outcome using diffusion-weighted images. Images obtained at the acute stage using a support vector machine, in which prediction is based on a set of input data, showed a correlation between poor motor outcome and changes in the corticospinal bundle and white matter tracts derived from the premotor cortex. Animal studies were used to compare different machine learning systems including generalized linear and additive models, support vector machine, adaptive boosting, and random forest. Such studies revealed that imaging-based algorithms can be used to estimate potentially remediable tissue depending on the variation in lesions following acute ischemic stroke.[4]

John Kalafut and colleagues at GE Healthcare Digital Solutions in collaboration with the MGH & BWH Center for Clinical Data Science (CCDS) are also developing a core assessment algorithm based on diffusion-weighted imaging, with exquisite sensitivity and specificity.

User-friendly Clinical Application

Despite the complexity of the deep learning networks and the steep learning curve involved, clinical application has been driven by user-friendly programs. For example, RAPID and PRONTO represent programs designed to provide clinicians easy access to machine learning. Compared with interfaces such as Keras or Lasagne that require knowledge of programming languages including Python, the nVidia’s DIGITS framework provides access to deep learning tools without the need to learn programming skills.1 Machine learning models are relatively independent of the potential interactions between clinical features and related factors. They are cherished clinically not only for their prognostic and predictive value, but also for the identification of patients with a wide range of treatment indications, including aggressive endovascular interventions.


  1. Deep learning guided stroke management: a review of clinical applications. Journal of Neurointerventional Surgery. https://jnis.bmj.com/content/10/4/358.long doi: 10.1136/neurintsurg-2017-013355 Accessed on June 28, 2018
  2. Deep learning of tissue fate features in acute ischemic stroke. 2015 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5597003 Accessed on June 28, 2018.
  3. Prediction of Tissue Outcome and Assessment of Treatment Effect in Acute Ischemic Stroke Using Deep Learning. Stroke. http://stroke.ahajournals.org/content/early/2018/05/01/STROKEAHA.117.019740. doi: 10.1161/STROKEAHA.117.019740 Accessed on June 28, 2018
  4. Machine Learning for Outcome Prediction of Acute Ischemic Stroke Post Intra-Arterial Therapy. PLoS ONE. http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0088225. doi:10.1371/journal.pone.0088225. Accessed on June 28, 2018