Evaluation of Diffusion Lesion Volume Measurements in Acute Ischemic Stroke Using Encoder-Decoder Convolutional...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Automatic segmentation of cerebral infarction using diffusion-weighted imaging (DWI) often relies on fixed apparent diffusion coefficient (ADC) thresholds.
• Fixed ADC threshold accuracy is limited by temporal variations in ADC values post-stroke onset.
• Deep learning offers potential for improved accuracy with sufficient annotated lesion data.

Purpose of the Study:

• Evaluate deep learning-based segmentation methods for cerebral infarction.
• Compare deep learning methods against commercial software (RAPID) for lesion volume measurement accuracy.
• Assess the performance of U-net models trained on DWI and ADC data versus DWI data alone.

Main Methods:

• Developed two U-net convolutional neural network models: U-net (DWI+ADC) and U-net (DWI).
• Trained models on 296 subjects and validated externally on 134 subjects.
• Used expert neurologist manual delineations as ground-truth and compared with RAPID analysis.

Main Results:

• U-net (DWI+ADC) achieved the highest intraclass correlation coefficient (1.0) with manual segmentation in external validation.
• U-net (DWI+ADC) demonstrated high correlation (0.99) with RAPID analysis results.
• Smallest Bland-Altman limits of agreement (-5.31 to 4.93 mL) were observed between U-net (DWI+ADC) and manual segmentation.

Conclusions:

• The developed deep learning method provides fully automatic and accurate segmentation of diffusion lesions.
• The method shows strong agreement with manual segmentation and commercial software for lesion volume quantification.
• This automated approach holds promise for patient selection in late-window endovascular reperfusion therapy for acute stroke.