Lateral Ventricle MR Image Segmentation Based on Recursive Gated Convolution with Multi-Scale Feature Fusion
Jinlong Wang, Xiufeng Zhang, Chenglong Wang · 2023
The lateral ventricle, a crucial component of the brain structure, holds immense significance in medical image analysis and clinical diagnosis. Nevertheless, accurately segmenting lateral ventricles remains a challenging endeavor due to variations in their shapes and sizes, as well as image noise. In response to this challenge, we propose a novel deep learning network architecture that combines recursive gated convolution and multi-scale feature fusion strategies. Recursive gated convolution, as an effective convolution operation, adeptly captures long-range dependencies in images, thereby enhancing the network's capacity to discern lateral ventricular regions. Concurrently, we introduce multi-scale feature fusion to improve the network's characterization ability and segmentation accuracy by integrating feature information from different scales. Subsequently, we conducted an extensive experimental evaluation utilizing the publicly available IBSR dataset. The results of our experiments demonstrate the superiority of our method in MRI lateral ventricle segmentation. Our model achieved outstanding Pre, Dice, and Jac results of 92.93%, 92.84%, and 89.39%, respectively. Compared to other algorithms, our method accurately localizes and segments the lateral ventricle region, yielding finer segmentation outcomes.