Leveraging anatomical constraints with uncertainty for pneumothorax segmentation.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Thoracic radiology

Background:

• Pneumothorax, an emergency condition, involves air in the pleural space, often overlooked by current deep learning (DL) models due to their end-to-end approach.
• Existing DL models for pneumothorax segmentation neglect the critical location-specific nature of the condition within the "lung + space" area on chest radiographs.

Purpose of the Study:

• To develop a novel deep learning approach for precise pneumothorax segmentation by incorporating the "lung + space" as a constraint.
• To enhance the reliability and accuracy of DL models in identifying pneumothorax by integrating domain knowledge.

Main Methods:

• Proposed a DL method using "lung + space" as a constraint for pneumothorax segmentation, trained with auxiliary lung segmentation tasks on external datasets.
• Implemented a discriminator to mitigate domain shift issues between auxiliary and target datasets, ensuring constraint reliability.
• Validated the approach across multiple DL architectures (U-Net, LinkNet, PSPNet) and backbones (VGG-11, MobileOne-S0).

Main Results:

• Achieved average performance improvements of 4.6% (IoU), 3.6% (DSC), and 3.3% (Hausdorff distance) across baseline models.
• Demonstrated consistent performance gains across different DL architectures and backbones.
• Ablation and robustness studies confirmed the effectiveness and stability of the proposed method's components.

Conclusions:

• Integrating medical domain knowledge, specifically the location-specific nature of pneumothorax, significantly enhances DL-based segmentation accuracy.
• The proposed method increases clinician trust in AI tools for medical diagnostics.
• This approach holds promise for improving segmentation of other thoracic conditions with location-relevant characteristics.