Multi-stage semi-supervised learning enhances white matter hyperintensity segmentation.

Area of Science:

• Medical imaging and artificial intelligence
• Neuroscience and neurology
• Machine learning for healthcare

Background:

• White matter hyperintensities (WMHs) are common in older adults and linked to increased dementia and stroke risk.
• Manual segmentation of WMHs is time-consuming and labor-intensive, hindering the creation of large annotated datasets.
• Limited annotated data presents a significant challenge for developing supervised machine learning models for WMH segmentation.

Purpose of the Study:

• To develop and evaluate a multi-stage semi-supervised learning (M3SL) approach for automated white matter hyperintensity (WMH) segmentation.
• To address the challenge of limited annotated data in WMH segmentation using a combination of un-annotated and gold-standard annotated data.
• To improve the generalization of WMH segmentation models across different MRI scanner vendors and clinical populations (cognitively normal, mild cognitive impairment, Alzheimer's disease).

Main Methods:

• Implemented a multi-stage semi-supervised learning (M3SL) framework integrating traditional processing methods for un-annotated data ('bronze' and 'silver') with a small set of 'gold'-standard annotations.
• Utilized the M3SL approach to fine-tune model weights within a U-Net architecture for WMH segmentation.
• Trained and validated the model using data from multiple scanners across three vendors and diverse clinical cohorts (cognitively normal, MCI, AD).

Main Results:

• The M3SL approach demonstrated superior generalization performance across different scanner vendors and clinical stages (CN, MCI, AD) compared to conventional and transfer-learning deep learning methods.
• Significant improvements in WMH segmentation accuracy were observed using M3SL, as measured by F-measure, IoU, and Hausdorff distance (p < 0.001).
• The study confirmed the utility of automated, non-machine learning tools within a multi-stage learning framework to enhance model performance.

Conclusions:

• The M3SL approach effectively overcomes the limitations of scarce annotated data for WMH segmentation.
• This methodology enhances model performance and generalization, offering a viable solution for large-scale WMH analysis in clinical research.
• Automated tools play a crucial role in semi-supervised learning frameworks, improving the efficiency and accuracy of medical image segmentation.