BBox-Guided Segmentor: Leveraging expert knowledge for accurate stroke lesion segmentation using weakly supervised...

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Neurology

Background:

• Stroke is a major cause of death and disability globally.
• Accurate stroke lesion segmentation is crucial for diagnosis and treatment.
• Current deep learning models struggle with limited labeled data and detecting small lesions.

Purpose of the Study:

• To develop an accurate stroke lesion segmentation method using weakly-supervised learning.
• To leverage expert-provided bounding box labels to improve segmentation accuracy.
• To address the challenge of insufficient labeled data in deep learning for stroke imaging.

Main Methods:

• Proposed a BBox-Guided Segmentor leveraging coarse bounding box labels.
• Employed a weakly-supervised approach with both bounding box and full segmentation labels.
• Utilized adversarial training with a large dataset of weakly labeled images.

Main Results:

• Achieved superior performance over state-of-the-art stroke lesion segmentation models.
• Demonstrated competitive performance as a fully supervised method using significantly less labeled data.
• Validated on a unique clinical dataset of 99 fully labeled and 831 weakly labeled cases.

Conclusions:

• The BBox-Guided Segmentor significantly improves stroke lesion segmentation accuracy.
• Weakly-supervised learning with bounding boxes is effective for addressing data scarcity.
• The method has the potential to enhance stroke diagnosis, treatment planning, and patient outcomes.