Pyramid feature adaptation for semi-supervised cardiac bi-ventricle segmentation

Area of Science:

• Medical Imaging Analysis
• Cardiovascular Research
• Artificial Intelligence in Medicine

Background:

• Cardiac bi-ventricle segmentation (BVS) is crucial for evaluating cardiac function (e.g., ejection fraction).
• BVS is challenging due to structural variability and limited labeled data in medical datasets.
• Accurate BVS is essential for reliable assessment of left ventricle (LV) and right ventricle (RV) volumes.

Purpose of the Study:

• To develop an effective semi-supervised method for cardiac bi-ventricle segmentation (BVS).
• To address the challenges of high structural variability and data scarcity in BVS.
• To improve the accuracy and clinical applicability of cardiac segmentation techniques.

Main Methods:

• Proposed a pyramid feature adaptation based semi-supervised method (PABVS) for BVS.
• Extracted multiscale pyramid features to handle bi-ventricle structure variability.
• Employed a weighted pyramid feature adaptation strategy using adversarial learning for smooth feature spaces between labeled and unlabeled data.

Main Results:

• Achieved a Dice score of 0.915 for EpiLV segmentation using only 40% labeled data.
• Reached a Dice score of 0.976 for EpiLV segmentation with fully labeled data.
• Outperformed existing mainstream semi-supervised methods in cardiac bi-ventricle segmentation.

Conclusions:

• The proposed PABVS method demonstrates high efficacy in cardiac bi-ventricle segmentation, even with limited labeled data.
• PABVS effectively handles structural variability through multiscale feature extraction and adaptation.
• The method shows significant potential for practical clinical applications in cardiac imaging.