Magnetic resonance imaging-based deep learning imaging biomarker for predicting functional outcomes after acute...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Neurology and Stroke Research

Background:

• Clinical risk scores are vital for predicting outcomes in acute ischemic stroke (AIS) patients.
• Deep learning (DL) offers advanced methods for developing predictive applications using magnetic resonance (MR) imaging.
• There is a need for improved biomarkers to enhance stroke outcome prediction.

Purpose of the Study:

• To develop a novel MR-based DL imaging biomarker for predicting functional outcomes in AIS patients.
• To evaluate the added predictive value of this DL biomarker to existing clinical risk scores.

Main Methods:

• A DL model was trained on MR images and radiomics data from 3338 AIS patients.
• The model generated a DL score, serving as an imaging biomarker for predicting 3-month functional outcomes.
• The biomarker's performance was compared against five established risk scores, and its incremental value was assessed.

Main Results:

• The DL imaging biomarker demonstrated strong predictive performance (AUC 0.788), comparable to several existing risk scores.
• Incorporating the DL biomarker significantly improved the predictive accuracy of the clinical risk scores (AUCs increased, p < 0.001).
• Net reclassification and integrated discrimination improvements confirmed the substantial added value of the DL biomarker.

Conclusions:

• Developing MR-based imaging biomarkers using DL is feasible for stroke outcome prediction.
• This DL imaging biomarker effectively enhances the predictive capabilities of current clinical risk scores.
• The findings support the integration of AI-driven imaging biomarkers into routine stroke care for improved patient management.