Iterative Segmentation from Limited Training Data: Applications to Congenital Heart Disease.

Area of Science:

• Medical image analysis
• Deep learning for medical imaging

Background:

• Accurate segmentation of cardiac structures in MRI is crucial for diagnosing congenital heart disease (CHD).
• Traditional methods often require large annotated datasets, which are difficult to obtain for rare conditions like CHD with diverse anatomical variations.

Purpose of the Study:

• To develop and evaluate a novel iterative segmentation model for cardiac MRI.
• To improve the accuracy of segmenting heart structures in patients with congenital heart disease (CHD).

Main Methods:

• Implemented a recurrent neural network (RNN) for iterative image segmentation.
• Trained the model by optimizing intermediate segmentation steps and the final output.
• Utilized incomplete or inaccurate input segmentations paired with recommended next steps during training.

Main Results:

• The iterative segmentation model was successfully trained on a small dataset (20 CHD patient images).
• The model accurately segmented individual heart chambers and great vessels.
• The iterative approach demonstrated superior accuracy compared to direct segmentation, particularly for severe CHD malformations.

Conclusions:

• The proposed iterative segmentation model offers an effective solution for segmenting cardiac MRI in CHD patients, even with limited data.
• This method alleviates challenges posed by anatomical variability and topological changes in CHD.
• The RNN-based iterative approach shows significant potential for clinical application in CHD diagnosis and management.