Introduction
The hippocampus is a critical subcortical structure involved in memory and learning. In clinical practice, precise segmentation of the hippocampus is essential for applications ranging from neurodegenerative disease diagnosis (e.g., Alzheimer’s disease and epilepsy) to precision radiotherapy. In whole-brain radiation therapy, sparing the hippocampus is a priority to mitigate cognitive decline. However, manual segmentation is labor-intensive and prone to inter-observer variability. Automated deep learning methods, particularly the 3D-UNet, have shown promise but often struggle with the hippocampus's small volume, low contrast, and low signal-to-noise ratio in MR images, frequently resulting in anatomically impossible "hollow" or fragmented segmentations.
Methodology
To address these topological inconsistencies, this study proposes an improved 3D-UNet architecture. The model utilizes a standard encoder-decoder backbone with skip connections but introduces a novel filling technique as a post-processing step to ensure structural integrity.
The Filling Technique
The core innovation is a maximum connected region algorithm designed to eliminate fragmentation. The algorithm evaluates predicted regions using a connectivity threshold $\theta$ based on the formula: $$(x_2 - x_1 + y_2 - y_1 + 2) \leq \theta$$ If two predicted regions fall within this threshold, they are merged. This effectively "fills" internal holes and connects fragmented pieces that are anatomically part of the same structure.
Training and Optimization
To compensate for the limited dataset and the class imbalance inherent in segmenting a small structure like the hippocampus, the researchers employed:
- Data Augmentation: Rotation, scaling, grey-value shifts, and smooth dense deformation fields were used to simulate anatomical diversity.
- Weighted Loss Function: A weighted cross-entropy loss was implemented to prioritize the small hippocampal volume over the dominant background pixels.
Experimental Setup
The model was evaluated using a dataset of 200 3D T1-weighted MR images from Hangzhou Cancer Hospital (2020–2022). The ground truth was established by three deputy chief physicians following RTOG 0933 clinical guidelines. The data was split into training (145), validation (30), and testing (25) sets.
Results
The integration of the filling technique yielded significant quantitative improvements, particularly at lower input resolutions:
- Patch-Based Performance: On 48x48x48 image patches, the proposed model achieved a Dice score of 0.7989, a dramatic increase from the 0.5457 achieved by the baseline 3D-UNet.
- High-Resolution Accuracy: Increasing the input size to 96x96x96 further improved the Dice score to 0.8674 and the mean Intersection over Union (mIoU) to 0.7668.
- Boundary Accuracy: The model showed a significant reduction in Hausdorff Distance (HD), indicating cleaner and more precise anatomical boundaries compared to standard architectures like VNet or SegResNet.
Conclusion
By incorporating a geometric filling technique, this improved 3D-UNet model successfully overcomes the common pitfalls of automated hippocampus segmentation. The approach ensures topological consistency, significantly reducing the need for manual correction by clinicians. While future work should focus on multi-center validation and larger datasets, this model represents a robust advancement for automated tools in neuro-oncology and neurology.
Summary updated