Abstract
Accurate segmentation of liver tumors from Computed Tomography (CT) scans is a critical yet challenging task in digital pathology due to the high variability in tumor size, shape, and the low contrast between lesions and healthy liver tissue. We present the Decoupled Pyramid Correlation Network (DPC-Net), a novel 3D medical imaging framework designed to address these challenges by explicitly decoupling spatial and semantic feature correlations. By utilizing two specialized attention modules—the Spatial Correlation (SpaCor) module for boundary refinement and the Semantic Correlation (SemCor) module for noise suppression—DPC-Net effectively bridges the gap between low-level textures and high-level anatomical context. Our evaluation on the MICCAI 2017 LiTS dataset demonstrates that DPC-Net achieves state-of-the-art performance, outperforming established models such as nnU-Net and H-DenseUNet.
1. Introduction
Liver cancer remains a leading cause of cancer-related mortality worldwide, making precise tumor segmentation essential for surgical planning and treatment monitoring. However, manual segmentation is time-consuming and prone to inter-observer variability. While deep learning models, particularly Fully Convolutional Networks (FCNs) like U-Net, have shown promise, they often struggle with the "unclear boundaries" characteristic of liver tumors.
Standard architectures typically rely on skip connections to merge low-level spatial features (edges and textures) with high-level semantic features (object identity). However, naive concatenation often introduces noise from shallow layers or fails to leverage the global context provided by deeper layers. The core limitation lies in the failure to account for the distinct roles that spatial and semantic dimensions play in feature correlation. DPC-Net addresses this by decoupling these correlations into specialized pathways.
2. Methodology
The DPC-Net architecture is built upon a 3D U-Net variant and introduces a hierarchical approach to feature refinement through three primary components:
2.1 Pyramid Feature Encoder (PFE)
The PFE serves as the backbone, extracting multi-level features from 3D CT volumes. To accommodate the small batch sizes inherent in high-resolution 3D imaging, the encoder replaces standard Batch Normalization with Instance Normalization, ensuring stable training and better generalization across varying scan intensities.
2.2 Spatial Correlation (SpaCor) Module
The SpaCor module operates as a top-down guidance system. It utilizes high-level semantic features to identify the general region of interest and then applies this context to low-level features to "carve out" precise boundaries. By generating a spatial attention map, SpaCor focuses the network's attention on the fine details of tumor edges, which are often lost in standard downsampling processes.
2.3 Semantic Correlation (SemCor) Module
Conversely, the SemCor module acts as a bottom-up filter. It leverages the rich category information found in deep layers to weigh the importance of low-level features. This mechanism allows the network to suppress irrelevant "distractors"—such as blood vessels or adjacent organs—that may share similar textural properties with tumors in shallow layers but lack the semantic identity of a lesion.
2.4 Training Objectives
The network is optimized using a hybrid loss function combining Cross-Entropy and Dice loss. Furthermore, the model employs deep supervision, where auxiliary losses are calculated at every level of the feature pyramid. This ensures that the intermediate layers are explicitly guided toward meaningful feature representation throughout the training process.
3. Experimental Setup
3.1 Dataset
The model was rigorously evaluated using the MICCAI 2017 LiTS (Liver Tumor Segmentation) Challenge dataset. This benchmark includes 131 contrast-enhanced 3D abdominal CT scans for training and 70 scans for testing, with ground truths for the test set held out by the challenge organizers.
3.2 Preprocessing
To ensure consistency across different scanners, the data underwent several preprocessing steps:
- Hounsfield Unit (HU) Clipping: Intensities were clipped to a specific range to highlight liver and tumor tissues.
- Normalization: Z-score normalization was applied to standardize the voxel intensity distribution.
- Resampling: Volumes were resampled to a fixed voxel spacing to maintain spatial consistency across the dataset.
4. Results and Discussion
DPC-Net demonstrated superior performance compared to several state-of-the-art methods on the LiTS testing set.
4.1 Segmentation Performance
The model achieved a Dice Similarity Coefficient (DSC) of 76.4% for tumor segmentation, notably outperforming the popular nnU-Net (74.8%) and H-DenseUNet (72.2%). For liver segmentation, the model remained highly competitive with a DSC of 96.0%.
4.2 Boundary Accuracy
A key metric for clinical utility is the Average Symmetric Surface Distance (ASSD). DPC-Net reduced the tumor ASSD to 0.838 mm, indicating that the predicted boundaries are significantly closer to the ground truth than those produced by previous architectures. This improvement is directly attributable to the SpaCor module's ability to refine edge details.
4.3 Efficiency
Despite the inclusion of specialized attention modules, DPC-Net remains computationally efficient. The SpaCor and SemCor modules add only approximately 0.52M parameters to the baseline 3D U-Net, making the framework practical for memory-constrained 3D medical imaging environments.
5. Conclusion
The Decoupled Pyramid Correlation Network (DPC-Net) represents a significant advancement in 3D liver tumor segmentation. By separating spatial and semantic correlations, the model effectively manages the trade-off between local detail and global context. While the current study focuses on single-phase CT scans, future work will explore extending DPC-Net to multi-phase data and integrating Vision Transformer (ViT) components to further enhance long-range dependency modeling. The results on the LiTS benchmark suggest that DPC-Net is a robust candidate for integration into automated clinical diagnostic pipelines.
Summary updated