Optimizing Acute Stroke Segmentation on MRI Using Deep Learning: Self-Configuring Neural Networks Provide High...

Area of Science:

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

Background:

• Accurate infarct segmentation is crucial for managing ischemic stroke and predicting outcomes.
• The role of combining Diffusion-Weighted Imaging (DWI), ADC, and FLAIR MRI sequences for deep learning-based infarct segmentation remains unclear.
• Automated model optimization through self-configuration technologies promises enhanced performance and generalizability.

Purpose of the Study:

• To assess the utility of DWI, ADC, and FLAIR MRI sequences for ischemic stroke segmentation using deep learning.
• To compare the performance of self-configuring nnU-Net models against conventional U-Net models.
• To evaluate the generalizability of the best-performing model on an external clinical dataset.

Main Methods:

• Trained 3D self-configuring nnU-Net and standard 3D U-Net models using MONAI on 200 infarcts.
• Utilized DWI, ADC, and FLAIR sequences individually and in combination.
• Compared segmentation results using a paired t-test on a 50-case hold-out set and validated externally on 50 MRIs.

Main Results:

• nnU-Net with DWI sequences achieved a Dice score of 0.810 ± 0.155.
• Supplementing DWI with ADC and FLAIR sequences did not yield statistically significant improvements (Dice score 0.813 ± 0.150, p=0.15).
• nnU-Net models significantly outperformed standard U-Net models across all sequence combinations (p < 0.001).
• External validation on 50 MRIs yielded a Dice score of 0.704 ± 0.199 for positive cases, with some false positives in cases of intracranial hemorrhage.

Conclusions:

• Highly optimized neural networks, such as nnU-Net, provide excellent stroke segmentation performance using only DWI images.
• There is no significant performance gain from adding ADC and FLAIR sequences when using nnU-Net.
• nnU-Net's superior performance and generalizability over standard U-Net architectures offer a strong foundation for optimized acute stroke segmentation in clinical MRI settings.