Efficient Segmentation for Left Atrium With Convolution Neural Network Based on Active Learning in Late Gadolinium...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac magnetic resonance imaging (CMRI) is crucial for diagnosing cardiovascular diseases.
• Accurate segmentation of cardiac structures, like the left atrium, is vital for quantitative analysis.
• Limited annotated datasets pose a challenge for developing robust segmentation models.

Purpose of the Study:

• To develop a fully automatic segmentation method for the left atrium in late gadolinium enhancement CMRI (LGE-CMRI).
• To leverage active learning to train segmentation models efficiently with limited datasets.
• To reduce the manual annotation time required for LGE-CMRI segmentation.

Main Methods:

• An active learning framework was implemented for left atrium segmentation in LGE-CMRI.
• A cohort of 98 patients with atrial fibrillation was enrolled.
• The model was iteratively trained using expert-annotated and human-in-the-loop corrected data.

Main Results:

• The segmentation accuracy, measured by Dice coefficients, improved across three training steps, reaching 0.90 ± 0.02.
• Bias in Bland-Altman plots decreased significantly, with the final step showing 2.68% (-8.57-13.93).
• Annotation time was drastically reduced from 218 ± 31 seconds to approximately 36.56 ± 15 seconds per case.

Conclusions:

• Deep active learning effectively reduces annotation time for LGE-CMRI.
• This method enables efficient training of segmentation models even with limited annotated data.
• The proposed approach facilitates automated left atrium segmentation in cardiac MRI.