Segmentation-based cardiomegaly detection based on semi-supervised estimation of cardiothoracic ratio

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Clinical integration of neural networks is limited by model interpretability and generalization.
• Existing AI models often function as 'black boxes', hindering trust in clinical applications.
• Cardiomegaly detection in radiographs is a critical diagnostic task.

Purpose of the Study:

• To develop an effective and interpretable AI architecture for cardiomegaly detection.
• To address the 'black box' nature and generalization issues of AI models in clinical radiology.
• To reduce manual annotation costs through semi-supervised learning.

Main Methods:

• Developed a two-stage neural network architecture for segmenting cardiac and thoracic regions.
• Employed semi-supervised learning for optimizing segmentation models due to limited pixel-level data.
• Classified cardiomegaly based on the estimated cardiothoracic ratio from segmentations.
• Assessed model generalization using a cross-domain evaluation.

Main Results:

• The segmentation outputs enhanced the interpretability of the final cardiomegaly classification.
• Semi-supervised optimization significantly reduced the need for manual annotation.
• The architecture demonstrated robust generalization capabilities in a cross-domain setting.
• The developed model proved effective for cardiomegaly detection.

Conclusions:

• The proposed architecture offers an interpretable and generalizable solution for AI-driven cardiomegaly detection.
• Semi-supervised learning is effective for training segmentation models with limited labeled medical data.
• This approach facilitates the integration of AI tools in clinical radiology workflows.