Active learning strategy and hybrid training for infarct segmentation on diffusion MRI with a U-shaped network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate stroke lesion segmentation on diffusion MRI is crucial for patient care.
• Current neural network methods often suffer from high false positive rates, limiting clinical utility.

Purpose of the Study:

• To propose and evaluate a novel two-phase hybrid training strategy for 3D deconvolutional neural networks.
• To improve the accuracy and reduce false positives in stroke lesion segmentation on diffusion MRI.

Main Methods:

• A hybrid learning scheme involving a regular phase (whole MRI training) and a hybrid phase (alternating whole MRI and actively selected patches).
• Expert segmentation of infarcts on diffusion MRI from 929 patients, with data split into 60% training, 20% validation, and 20% testing sets.
• Comparison of segmentation performance between regular and hybrid training phases on the test population.

Main Results:

• Statistically significant improvement in Dice coefficient with hybrid training compared to regular training.
• Achieved a mean Dice score of [insert value from math tag].
• Reduced false positives by approximately 30% using the hybrid training approach.

Conclusions:

• The proposed hybrid training strategy enhances the performance of deep neural networks for stroke lesion segmentation.
• This approach leads to more accurate infarct segmentation and fewer false positives, benefiting clinical practice.
• The hybrid method represents a significant advancement in automated diffusion MRI analysis for stroke.