Abstract
In this paper, a novel 3D deep learning network is proposed for brain MR image segmentation with randomized connection, which can decrease the dependency between layers and increase the network capacity. The convolutional LSTM and 3D convolution are employed as network units to capture the long-term and short-term 3D properties respectively. To assemble these two kinds of spatial-temporal information and refine the deep learning outcomes, we further introduce an efficient graph-based node selection and label inference method. Experiments have been carried out on two publicly available databases and results demonstrate that the proposed method can obtain competitive performances as compared with other state-of-the-art methods.
1. Introduction
Segmenting sub-cortical structures (such as the Hippocampus) or pathological regions (like stroke lesions) in brain Magnetic Resonance (MR) images is a critical task for diagnosing neurodegenerative diseases and clinical planning. However, medical image segmentation faces several significant hurdles:
- Dimensionality and Spatial Dependencies: Medical scans are inherently 3D volumes. Standard 2D Convolutional Neural Networks (CNNs) process slices independently, losing crucial "depth" information. Conversely, while 3D CNNs capture volumetric features, they are computationally intensive and often limited to short-range spatial dependencies due to kernel size constraints.
- Low Contrast and Ambiguous Boundaries: Brain scans frequently exhibit poor contrast between adjacent tissues, making it difficult for models to delineate precise object boundaries.
- Computational Efficiency: Dense 3D prediction is slow. Traditional post-processing refinement methods, such as Fully Connected Conditional Random Fields (CRFs), become computationally prohibitive when processing millions of voxels in a 3D volume.
2. Methodology
The proposed framework addresses these challenges through a three-pronged approach: hybrid network units, a randomized connection architecture, and efficient graph-based refinement.
2.1 Hybrid Network Units
To capture both local and global spatial-temporal features, the network integrates two types of units:
- 3D Convolutions: These handle short-term, local spatial patterns within a small volumetric neighborhood.
- Convolutional LSTMs (ConvLSTM): By treating the third axis (the slices of the brain) as a temporal dimension, ConvLSTMs allow the model to maintain long-term dependencies across the entire volume, effectively solving the "short-sightedness" of standard 3D kernels without excessive memory overhead.
2.2 Randomized Connection Network
The architecture follows a symmetric contracting and expansive path, similar to the U-Net. However, the authors introduce Randomized Connections between these paths. During training, skip connections are activated with a probability $\alpha$. This stochasticity forces the network to behave as an ensemble of multiple sub-networks, which acts as a powerful regularizer—similar to dropout but at the architectural level—reducing layer dependency and increasing overall capacity.
2.3 Graph-based Label Inference
To avoid the heavy computational cost of standard CRFs, a two-step refinement process is introduced:
- Node Selection: The model first identifies "easy" voxels with high classification confidence and prunes them from the refinement stage.
- Label Inference: A Random Walker framework is applied only to the remaining "hard" candidate pixels, typically located near boundaries. This targeted approach significantly accelerates post-processing.
3. Experimental Results
The model was evaluated on two major benchmarks: the LPBA40 dataset (56 sub-cortical structures) and the ISLES Challenge 2016 (Ischemic Stroke Lesion Segmentation).
3.1 Performance Metrics
On the LPBA40 dataset, the proposed method outperformed standard FCNs and 3D U-Nets. The randomized connection mechanism improved the Dice Coefficient (DC) by 1.26% for 3D CNNs and 2.03% for ConvLSTMs. In the ISLES 2016 challenge, the team ranked 3rd among participants who completed all test cases, demonstrating robustness in segmenting irregular and unpredictably located lesions.
3.2 Computational Efficiency
The graph-based label inference proved highly efficient, taking only 4 seconds per structure compared to 60 seconds for a traditional fully connected CRF. This speedup is achieved by focusing computational power on the 0.1% of voxels that actually require refinement.
4. Conclusion
The 3D Randomized Connection Network offers a robust and efficient solution for volumetric brain MR image segmentation. By treating the Z-axis as a temporal sequence and employing randomized skip connections, the model achieves high accuracy with reduced layer dependency. While the initial node selection threshold remains a hyperparameter requiring tuning, the overall framework demonstrates state-of-the-art potential for clinical applications. Future work may focus on automating threshold selection or integrating the graph inference directly into end-to-end training.
Summary updated