A multimodal multitask deep learning model for predicting stroke lesion and functional outcomes using 4D CTP imaging and clinical metadata.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, University of Calgary, Calgary, Canada. [email protected].
- Hotchkiss Brain Institute, University of Calgary, Calgary, Canada. [email protected].
- Department of Radiology, University of Calgary, Calgary, Canada.
- Department of Diagnostic and Interventional Neuroradiology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany.
- Hotchkiss Brain Institute, University of Calgary, Calgary, Canada.
- Department of Clinical Neurosciences, University of Calgary, Calgary, Canada.
- Department of Community Health Sciences, University of Calgary, Calgary, Canada.
- Department of Radiology, University of Michigan, Ann Arbor, United States.
- Alberta Children's Hospital Research Institute, University of Calgary, Calgary, Canada.
Abstract
Acute ischemic stroke is a major global health challenge, leading to long-term disability or death without timely intervention. Among neuroimaging modalities, spatio-temporal (4D) computed tomography perfusion (CTP) is widely used to assess cerebral blood flow and guide acute treatment decisions. Beyond this role, recent studies have demonstrated its potential to predict lesion outcomes (irreversible tissue damage at follow-up) and functional outcomes (long-term functional independence). Although inherently related, these outcomes are typically modeled and predicted separately, ignoring shared patterns that could enhance predictive accuracy. Multitask learning provides a promising solution by leveraging shared representations across both related tasks. However, only a few stroke studies have adopted this approach so far, and those that did primarily relied on imaging data alone. Thus, we developed CTPredict, the first multimodal, multitask deep learning model that simultaneously predicts follow-up stroke lesions and functional outcomes (specifically, 90-day modified Rankin Scale) from 4D CTP imaging and clinical metadata. CTPredict integrates modality-specific encoders for feature extraction, a multimodal fusion module with cross-attention mechanisms to focus on relevant features from both data sources, and task-specific branches for outcome prediction, all within a computationally efficient framework for multitask learning. Evaluated on a challenging multi-center dataset of 111 AIS patients, CTPredict achieved a 0.23 Dice score and 0.77 accuracy for the lesion and functional outcome prediction tasks, respectively, outperforming single-task variants (0.21 Dice score, 0.73 accuracy). These results demonstrate the benefits of multitask learning and highlight CTPredict's potential to enable more streamlined, data-driven, and personalized stroke outcome predictions in clinical practice.