Deep learning-based automatic segmentation of cerebral infarcts on diffusion MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate segmentation of cerebral infarcts on MRI is crucial for diagnosis and treatment.
• Deep learning algorithms show promise but require large, diverse datasets for optimal performance.
• Challenges include data variability across different hospital sites and imaging protocols.

Purpose of the Study:

• To evaluate the impact of training data size and site diversity on deep learning models for infarct segmentation.
• To assess the effectiveness of cross-site domain adaptation techniques.
• To determine optimal strategies for enhancing algorithm generalizability and performance.

Main Methods:

• Utilized 10,820 annotated diffusion-weighted images (DWIs) from 10 university hospitals for training and internal testing.
• Trained 3D U-net based algorithms with varying sample sizes (217 to 8661 DWIs).
• Performed external validation on independent datasets and applied domain adaptation using subsets of external data.

Main Results:

• Increasing multi-site training data from 217 to 1732 DWIs significantly improved Dice similarity coefficient (DSC) and average Hausdorff distance (AHD).
• Further increases in data size yielded diminishing returns, with marginal performance gains.
• Domain adaptation with as few as 50 external images enabled a model trained on 217 images to perform comparably to one trained on 8661 images.

Conclusions:

• Multi-site training data, around 2000 DWIs, is essential for robust infarct segmentation.
• Cross-site domain adaptation is a highly effective strategy to improve performance and generalizability, especially with limited external data.
• Deep learning models benefit significantly from diverse, multi-site data and adaptation for reliable cerebral infarct segmentation.