Related Experiment Video
Updated: Mar 8, 2026
A Magnetic Resonance Imaging Protocol for Stroke Onset Time Estimation in Permanent Cerebral Ischemia
Published on: September 16, 2017
Mixture-of-Skip-Connection Deep Learning Model to Classify Stroke Severity from Diffusion Weighted Imaging Based on
Seunghun Oh1, Hyunsu Jeong1, Yoonjae Cho1
1Medical Science and Engineering, Graduate School of Artificial Intelligence (GSAI), Department of Electrical Engineering, Convergence IT Engineering, Mechanical Engineering and Medical Device Innovation Center, Pohang University of Science and Technology (POSTECH), 77 Cheongam-ro, Nam-gu, Pohang, 37673, Republic of Korea.
Abstract:
The National Institutes of Health Stroke Scale (NIHSS) is a quantitative tool, grading neurological deficits and guiding acute stroke management; however, bedside scoring is labor-intensive, time-consuming, and subject to interrater variability. This study proposes the TRansformer And Coordinate-aTtention NETwork (TRACT-NET), which includes a mixture-residual block, to predict stroke severity based on NIHSS score and diffusion-weighted imaging (DWI) scans. A 3D coordinate-attentive module and a 3D efficient self-attentive module were embedded into the mixture block for skip connection, enhancing feature representation by capturing axis awareness and long-range dependency. In the bottleneck stage, Mamba, a selective state-space model, was used to refine feature maps capturing relevant information from long sequences. TRACT-NET was evaluated with three-fold cross-validation of DWI and NIHSS data from 273 patients at Asan Medical Center (AMC) and externally validated on Stroke Outcome Optimization Project (SOOP) dataset, which includes 1106 patients. Patients were divided into binary groups: minor and non-minor strokes. TRACT-NET outperformed other classification models in predicting binary stroke severity. On the AMC dataset, it achieved 0.81 sensitivity, 0.9645 specificity, 0.8188 positive predictive value (PPV), 0.8054 negative predictive value (NPV), 0.7946 F1-score, 0.8137 accuracy, and 0.8137 area under the curve (AUC). On the SOOP dataset, the model achieved 0.6896 sensitivity, 0.7881 specificity, 0.7054 PPV, 0.6360 NPV, 0.6875 F1-score, 0.6896 accuracy, and 0.7094 AUC. Receiver operating characteristic curves demonstrated robust discriminative performance, whereas gradient-weighted class activation mapping results highlighted that TRACT-NET focused on clinically relevant DWI regions. These findings suggest TRACT-NET could assist stroke severity assessment in emergency situations, facilitating effective treatment decisions.
