Efficient Segmentation for Left Atrium With Convolution Neural Network Based on Active Learning in Late Gadolinium Enhancement Magnetic Resonance Imaging
Authors
Yongwon Cho, Hyungjoon Cho, Jaemin Shim, Jong-Il Choi, Young-Hoon Kim, Namkug Kim, Yu-Whan Oh, Sung Ho Hwang
This study proposes an automatic segmentation method for the left atrium using deep active learning in late gadolinium enhancement cardiac magnetic resonance imaging (LGE-CMRI). Ninety-eight atrial fibrillation patients from Korea University Anam Hospital were included. Initially, 20 cases were delineated by experts to create a draft model. This model was then refined using an additional 50 cases corrected through a human-in-the-loop process, culminating in a final training set of 98 cases. The segmentation's accuracy improved across three steps, with Dice coefficients of 0.85, 0.89, and 0.90, respectively. Biases in Bland-Altman plots decreased progressively, indicating improved consistency. Annotation times were significantly reduced from 218 seconds to approximately 36 seconds. This framework demonstrated that deep active learning significantly reduces annotation time and enables efficient training with a limited dataset in LGE-CMRI, making it a valuable approach for automating left atrium segmentation.