ICH-PRNet: a cross-modal intracerebral haemorrhage prognostic prediction method using joint-attention interaction...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neurology and Clinical Prognostics
• Deep Learning for Healthcare

Background:

• Accurate prediction of intracerebral hemorrhage (ICH) prognosis is vital for patient management.
• Current uni-modal deep learning methods show limitations due to ICH complexity.
• Existing cross-modal approaches struggle to effectively integrate diverse data types.

Purpose of the Study:

• To introduce ICH-PRNet, a novel cross-modal network for predicting ICH prognosis.
• To enhance the extraction of complementary information and cross-modal features.
• To improve the accuracy and efficacy of prognostic assessments in ICH patients.

Main Methods:

• Developed a joint-attention interaction encoder for integrating CT images and clinical texts.
• Implemented a multi-loss function to optimize cross-modal fusion.
• Utilized a self-adaptive dynamic prioritization algorithm for training balance.

Main Results:

• ICH-PRNet effectively integrates computed tomography images and clinical texts.
• The model establishes robust semantic connections and uncovers complementary cross-modal information.
• Demonstrated superior prediction results compared to state-of-the-art methods on multiple datasets.

Conclusions:

• The proposed ICH-PRNet model significantly advances ICH prognosis prediction.
• Effective cross-modal fusion of imaging and text data is key to improved outcomes.
• The novel approach offers a more accurate and efficient tool for neurosurgeons.