A multimodal Vision-Mamba model based on non-contrast CT hematoma and shell features predicts early hematoma expansion in hypertensive intracerebral hemorrhage: a multicenter study

@article{yu2026multimodal,
  title = {{A multimodal Vision-Mamba model based on non-contrast CT hematoma and shell features predicts early hematoma expansion in hypertensive intracerebral hemorrhage: a multicenter study}},
  author = {Jinna Yu and Lingting Kong and Yonggang Qiu and Yuanyuan Wang and Cong He and Lingfeng Fu and Chao Wang and Shisong Miao and Zigang Yuan and Lei Pei and Dong Xie},
  journal = {BMC Medical Imaging},
  year = {2026},
  month = aug,
  doi = {10.1186/s12880-026-02662-z},
  url = {https://doi.org/10.1186/s12880-026-02662-z},
  note = {CC BY-NC-ND 4.0},
}

TY  - JOUR
TI  - A multimodal Vision-Mamba model based on non-contrast CT hematoma and shell features predicts early hematoma expansion in hypertensive intracerebral hemorrhage: a multicenter study
AU  - Jinna Yu
AU  - Lingting Kong
AU  - Yonggang Qiu
AU  - Yuanyuan Wang
AU  - Cong He
AU  - Lingfeng Fu
AU  - Chao Wang
AU  - Shisong Miao
AU  - Zigang Yuan
AU  - Lei Pei
AU  - Dong Xie
PY  - 2026
DA  - 2026/08/01
JO  - BMC Medical Imaging
DO  - 10.1186/s12880-026-02662-z
UR  - https://doi.org/10.1186/s12880-026-02662-z
AB  - Objective To develop and evaluate a multimodal Vision-Mamba deep learning (DL) model integrating clinical data, noncontrast computed tomography (NCCT) hematoma features, and shell features to predict early hematoma expansion (EHE) in hypertensive intracerebral hemorrhage (HICH). Methods In this multicenter retrospective study, a total of 676 patients diagnosed with HICH who received baseline NCCT within 6 hours and subsequent imaging within 24 hours were included. Patients were categorized as EHE-positive and EHE-negative and divided by center into training (n = 361), validation (n = 188), and test (n = 127) sets. Hematoma regions were automatically segmented using No new U-Net (nnU-Net), and a perihematomal shell feature map was derived. Four Vision-Mamba models were constructed: hematoma, shell, hematoma_shell, and Combined model (a combined model integrating hematoma, shell, and clinical data). For comparison, we also constructed conventional DL architectures including DenseNet121, EfficientNet, ResNet18, and Transformer. All baseline comparison models shared an identical structure (hematoma and shell), incorporating clinical feature fusion. The assessment of model performance was conducted through the area under the receiver operating characteristic curve (AUC), along with DeLong’s test, calibration curves, and decision curve analyses. Results Within the test set, the AUCs for the hematoma, shell, and hematoma_shell models were recorded as 0.779, 0.812, and 0.832, respectively. The Combined model demonstrated the best performance in predicting EHE, achieving AUCs of 0.966 (95% CI: 0.949–0.980) in the training set, 0.938 (95% CI: 0.902–0.968) in the validation set, and 0.941 (95% CI: 0.899–0.973) in the test set. DeLong’s test in the test set (n = 127) confirmed that the Combined model (AUC = 0.941) was significantly superior to all DL baselines: DenseNet121 (AUC = 0.851, p = 0.018), EfficientNet (AUC = 0.769, p < 0.001), ResNet18 (AUC = 0.778, p < 0.001), and Transformer (AUC = 0.804, p < 0.001). Decision curve analysis indicated a higher net benefit than treat-all and treat-none strategies across an observed threshold range of approximately 0.35–0.80. Conclusion The multimodal Vision-Mamba DL model based on clinical variables, NCCT hematoma features, and shell features showed predictive performance for EHE risk stratification in patients with HICH. These findings provide preliminary evidence supporting the use of imaging biomarkers for EHE risk assessment.
SN  - 1471-2342
ER  -