Optimizing stroke lesion segmentation: A dual-approach using Gaussian mixture models and nnU-Net.

Area of Science:

• Biomedical imaging
• Machine learning
• Neurology

Background:

• Machine learning models are vital for stroke lesion segmentation in biomedical imaging.
• Gaussian Mixture Models (GMM) and nnU-Net are prominent segmentation workflows.
• Current evaluations often overlook model reliability in detecting therapy-induced changes.

Purpose of the Study:

• To systematically evaluate GMM and nnU-Net for detecting therapy-related changes in stroke volume.
• To assess the impact of different ground truth (GT) definitions on model performance in therapy assessment.

Main Methods:

• Comparative analysis of GMM and nnU-Net segmentation performance.
• Evaluation of nnU-Net trained on manual versus GMM-derived GT labels for therapy response detection.

Main Results:

• Both GMM and nnU-Net exhibit strong segmentation accuracy.
• nnU-Net trained solely on manual segmentations failed to detect significant therapy-induced stroke volume reductions.
• nnU-Net trained with GMM-derived GT labels demonstrated improved detection of therapy response.

Conclusions:

• Segmentation accuracy alone is insufficient for evaluating models in therapy assessment.
• The choice of ground truth definition significantly influences a model's ability to detect treatment effects.
• Integrating probabilistic methods with deep learning enhances therapy response detection in stroke imaging.