TMRKP: Trustworthy MRI-based radiology knowledge-aware prediction framework for stroke-associated pneumonia and hemorrhagic transformation.
Authors
Affiliations (5)
Affiliations (5)
- Faculty of Intelligence Technology, Shanghai Institute of Technology, Shanghai, 201418, China; Institute of Robotics and Automatic Information System, Nankai University, Tianjin, 300071, China.
- Department of Radiology, Zhongda Hospital, School of Medical, Southeast University, Nanjing, Jiangsu, 210009, China.
- Department of Radiology, Jiangdu People's Hospital, Yangzhou University, Yangzhou, Jiangsu, 225000, China.
- Department of Radiology, The First Affiliated Hospital, Wenzhou Medical University, Wenzhou, Zhejiang, 325035, China. Electronic address: [email protected].
- Faculty of Intelligence Technology, Shanghai Institute of Technology, Shanghai, 201418, China. Electronic address: [email protected].
Abstract
Prediction of stroke-associated pneumonia (SAP) and hemorrhagic transformation (HT) is critical for improving the prognosis of acute ischemic stroke patients. However, current brain magnetic resonance imaging (MRI)-based models narrowly focus on generic visual patterns, and frequently overlook the radiology-informed knowledge critical for predicting SAP and HT. To address this limitation, we propose a trustworthy MRI-based radiology knowledge-aware prediction (TMRKP) framework equipped with radiology knowledge-aware masked modeling (RKMM) and evidential multi-sequence fusion (EMF). Specifically, TMRKP first organizes four MRI sequences into a 2 × 2 super-sequence configuration, mimicking the diagnostic perspective of radiologists who simultaneously analyze multiple imaging modalities. Second, to introduce radiology knowledge into representation learning, we focus on white matter hyperintensities (WMH), a well-established biomarker of chronic small-vessel disease and long-term ischemic injury that is clinically relevant to SAP and HT. During pre-training, RKMM segments WMH masks from the MRI sequence to steer the model toward WMH regions, enabling it to learn meaningful representations of these clinically relevant pathological areas. Third, for SAP and HT prediction, EMF models the predictive uncertainty using evidence derived from both the super-sequence and the WMH-aware masked super-sequence, and fuses the resulting multi-sequence evidence based on Dempster-Shafer theory (DST). Extensive experiments on the collected Jiangdu dataset, together with independent external validation on the Zhongda dataset, demonstrate the superiority and generalizability of TMRKP over the state-of-the-art models.