Segmenting Small Stroke Lesions with Novel Labeling Strategies.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Deep neural networks (DNNs) excel at stroke lesion segmentation.
• Accurate segmentation of small lesions is crucial for stroke diagnosis but remains challenging.
• Existing DNNs struggle with precise delineation of minimal stroke damage.

Purpose of the Study:

• To introduce novel methods for improving small stroke lesion segmentation.
• To enhance the accuracy of DNNs in identifying and outlining small brain lesions post-stroke.
• To provide adaptable techniques for various neural network architectures.

Main Methods:

• Proposed Multi-Size Labeling (MSL) by categorizing lesion masks based on volume.
• Introduced Distance-Based Labeling (DBL) to emphasize lesion boundaries.
• Integrated MSL and DBL into existing DNN frameworks for evaluation.

Main Results:

• An ensemble of MSL and DBL outperformed the 2022 MICCAI ATLAS Challenge winner in recall, F1, and Dice scores.
• MSL and DBL achieved superior or equal performance on both small lesion subsets and the entire dataset.
• A single MSL model surpassed prior ensemble methods on the mini-lesion subset, improving F1 and Dice scores.

Conclusions:

• MSL and DBL are effective strategies for enhancing small stroke lesion segmentation.
• These methods offer significant improvements over existing techniques, particularly for minimal lesions.
• The proposed approaches are readily integrable and boost the diagnostic capabilities of AI in stroke care.