StrokeDiffNet: quantifying DWI-FLAIR mismatch via a common feature space for time since stroke classification

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Time since stroke onset (TSS) is critical for acute ischemic stroke (AIS) treatment.
• Current DWI-FLAIR mismatch methods struggle with modality differences and feature alignment for precise TSS estimation.
• Accurate TSS quantification is essential for effective clinical decision-making in AIS.

Purpose of the Study:

• To develop a novel deep learning model, StrokeDiffNet, for accurate TSS estimation in AIS.
• To address the challenges of modality-style differences and feature misalignment in DWI and FLAIR imaging.
• To improve the quantitative measurement of inter-modality differences for enhanced TSS classification.

Main Methods:

• Proposed StrokeDiffNet to map DWI and FLAIR features into a common feature space (CFS).
• Employed self-reconstruction and interactive supervision for DWI-FLAIR encoder training to promote feature alignment.
• Introduced a cross-domain mixed-noise strategy for self-supervised training and a key-feature alignment strategy.

Main Results:

• StrokeDiffNet achieved strong TSS classification performance.
• Achieved accuracy of 73.0%, precision of 78.3%, F1-score of 78.3%, and AUC of 71.3%.
• Outperformed other mainstream classification networks in TSS classification tasks.

Conclusions:

• StrokeDiffNet effectively overcomes modality-style interference and improves key feature alignment for TSS estimation.
• The proposed method offers a promising approach for quantitative mismatch measurement and TSS classification in AIS.
• StrokeDiffNet demonstrates significant potential for improving clinical decision-making in acute ischemic stroke management.