MDIFN: a deep learning-based multi-dimensional interaction fusion network for placental MRI segmentation.
Authors
Affiliations (6)
Affiliations (6)
- Hainan General Hospital (Hainan Affiliated Hospital of Hainan Medical University), Haikou, China.
- Hainan Women and Children's Medical Center, Haikou, China.
- Nanophotonics Research Center, Institute of Microscale Optoelectronics & State Key Laboratory of Radio Frequency Heterogeneous, Shenzhen University, Shenzhen, China.
- Big Health Informatics Research Center, Fudan University, Shanghai, China.
- School of Artificial Intelligence Ningbo University of Technology, Ningbo, China.
- The Third People's Hospital of Longgang, Clinical Institute of Shantou University, Medical College (The Third People's Hospital of Longgang District Shenzhen), Shenzhen, China.
Abstract
The placenta serves as a vital channel for maternal-fetal substance exchange and is a crucial organ for maintaining pregnancy. Placental diseases can severely affect maternal health and fetal growth. Given the limitations of single-dimension segmentation models and traditional fusion strategies in meeting the demands of automatic placental magnetic resonance imaging (MRI) segmentation, this study developed an automatic placental MRI segmentation method based on a multi-dimensional interaction fusion network (MDIFN). This study included 267 placental MRI cases that met the inclusion criteria. After preprocessing steps, such as bias field correction, resampling, intensity normalization, and dimension matching, a dual-branch foundational architecture combining two-dimensional UNet (2D-UNet) and three-dimensional UNet (3D-UNet) was constructed. An interaction fusion bidirectional gated cross-dimension module was designed to achieve layer-wise synergistic fusion of two-dimensional (2D) detail features and three-dimensional (3D) contextual features. A global context recalibration module and a residual refinement module were introduced to optimize feature quality. A homoscedastic uncertainty-based multi-task weighted loss function was employed to balance the training process. Model performance was compared against six mainstream models using 10-fold cross-validation, and ablation studies were conducted to validate the contribution of core modules. The model achieved a Dice coefficient of 0.8749±0.0132 (P<0.05) on the test set, representing an improvement of 16.2% compared to pure 2D-UNet and 8.4% compared to pure 3D-UNet. The sensitivity was 0.8999±0.0172 (P<0.05), and the F1 score was 0.8749±0.0132, which were the optimal values among all models. The relative volume difference was 0.0301±0.0173, and the 95% Hausdorff distance was 7.67±1.04. Ablation experiments showed that removing the gating mechanism, global context module, or bidirectional interaction reduced the Dice coefficient to 0.8554, 0.8551, and 0.8566, respectively, confirming that each component contributes to performance improvement. This study addresses the limitations of single-dimension segmentation and traditional fusion models, providing an efficient and objective technical approach for automatic placental segmentation.