Automatic Prediction of Stroke Treatment Outcomes: Latest Advances and Perspectives

TL;DR

This paper surveys deep learning approaches for predicting stroke treatment outcomes, focusing on final infarct prediction and functional outcomes (mRS) using imaging and multimodal data. It reviews datasets such as ISLES 2017 and multi-center trials, along with outcome metrics like mRS and TICI, and model families including U-Net variants, 3D CNNs, and transformers, including 4D temporal networks. It highlights that multimodal fusion and incorporation of final infarct information improve predictive accuracy, with reported AUCs up to around 0.92 and Dice scores in the 0.3–0.6 range depending on task. It also notes challenges, including limited public benchmarks, data heterogeneity, sparse external validation, and interpretability concerns, and proposes directions such as adaptive fusion, federated learning, and annotation-free segmentation to accelerate clinical translation.

Abstract

Stroke is a major global health problem that causes mortality and morbidity. Predicting the outcomes of stroke intervention can facilitate clinical decision-making and improve patient care. Engaging and developing deep learning techniques can help to analyse large and diverse medical data, including brain scans, medical reports and other sensor information, such as EEG, ECG, EMG and so on. Despite the common data standardisation challenge within medical image analysis domain, the future of deep learning in stroke outcome prediction lie in using multimodal information, including final infarct data, to achieve better prediction of long-term functional outcomes. This article provides a broad review of recent advances and applications of deep learning in the prediction of stroke outcomes, including (i) the data and models used, (ii) the prediction tasks and measures of success, (iii) the current challenges and limitations, and (iv) future directions and potential benefits. This comprehensive review aims to provide researchers, clinicians, and policy makers with an up-to-date understanding of this rapidly evolving and promising field.

Figures (7)

• Figure 1: Analysis of stroke time frames and the predicted tasks at each time point. These predictions are based on initial scans performed upon hospital admission to aid physicians in their clinical decision-making process.
• Figure 2: The multimodal architecture introduced in robben2020prediction. The network exploits imaging and clinical information to predict the follow-up stroke lesion. : computed tomography perfusion, AIF: arterial input function which refers to the measure of blood concentration that enters the affected brain tissue after a stroke, Metadata: clinical information e.g. time to scanning, recanalization, completeness of recanalization. Yellow and blue rectangles represent patches extracted from scan, purple rectangle represents AIF in the volume of interest and red rectangle is the target region of interest to segment lesion on follow-up scan. Figure from robben2020prediction.
• Figure 3: The overview of the temporal U-Net proposed by amador2022predicting to estimate the final lesion mask of stroke. Each scan is fed separately into an encoder, and the temporal convolutional block encodes temporal information and pass the features to the decoder for the purpose of predicting the final lesion mask. Figure from amador2022predicting.
• Figure 4: Sample results of Gutierrez_2024 for three patients from each treatment group of (a) thrombolysis (IVT) and (b) thrombectomy (IA). Five columns for each patient: (1) the original follow-up image, (2) the baseline image average, (3) the prediction for the patient's treatment group, (4) the prediction from the alternative treatment model, and (5) a difference map highlighting changes in intensity between predictions. The ground truth lesion mask is outlined in white within the difference map, with orange indicating increased intensity in the alternative treatment prediction (column 4) and blue indicating decreased intensity. Figure from Gutierrez_2024.
• Figure 5: The architecture proposed by pinto2021combining to predict stroke outcome. The first part of the network includes unsupervised training that uses two rbm and the second part of the network comprises of and gated-s for supervised training. Figure from pinto2021combining.