Deep learning-based temporal muscle quantification on MRI predicts adverse outcomes in acute ischemic stroke

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurology and stroke research

Background:

• Acute ischemic stroke (AIS) poses significant health challenges.
• Accurate prognostic markers are crucial for patient management and rehabilitation planning.
• Temporal muscle (TM) measurements may offer prognostic insights in AIS.

Purpose of the Study:

• To develop a deep learning (DL) pipeline for automated analysis of temporal muscle (TM) parameters in AIS patients.
• To assess the prognostic value of TM thickness (TMT) and TM area (TMA) in predicting 6-month outcomes after AIS.
• To enable rapid and scalable quantification of TMT and TMA for clinical integration.

Main Methods:

• A DL pipeline was developed using ResNet50 for slice selection and TransUNet for TM segmentation.
• The pipeline was trained and validated on datasets comprising 1020 AIS patients.
• Performance was evaluated using accuracy, ±1 slice accuracy, mean absolute error, and Dice similarity coefficient (DSC).
• The association between DL-quantified TMT and TMA and 6-month outcomes was analyzed.

Main Results:

• The DL system achieved high accuracy in slice selection (72.91%) and TM segmentation (DSC of 0.858).
• Automatically quantified TMT and TMA were independently associated with poor 6-month outcomes in AIS patients.
• Hazard ratios indicated a protective effect of higher TMT and TMA on outcomes (TMT: 0.736, TMA: 0.702).

Conclusions:

• Temporal muscle thickness (TMT) and area (TMA) are significant prognostic markers in AIS.
• The developed end-to-end DL pipeline provides automated, rapid quantification of TMT and TMA.
• This technology supports scalable risk stratification and personalized rehabilitation planning for AIS patients.