Abstract
Nasopharyngeal carcinoma (NPC) is a malignant tumor requiring accurate and automatic segmentation of organs at risk (OAR) in computed tomography (CT) images for clinical treatment. While deep learning models like U-Net are common, they often struggle with variable OAR sizes and small organs due to insufficient use of global and multi-size information. This paper proposes SECP-Net, which utilizes SE-connection (SEC) modules and a pyramid structure to extract multi-scale information. Additionally, an auto-context cascaded network is designed to refine segmentation. Experimental results on head and neck CT images demonstrate that SECP-Net achieves state-of-the-art performance in Dice and Jaccard metrics, particularly for small anatomical structures.
Introduction
In radiotherapy for Nasopharyngeal Carcinoma (NPC), clinicians must manually delineate "Organs At Risk" (OAR)—such as the eyes, brain stem, and thyroid—to prevent radiation damage to healthy tissue. This manual process is labor-intensive and prone to inter-observer variability. While deep learning architectures like U-Net have automated parts of this workflow, they face significant challenges in this specific domain:
- Noisy Skip Connections: Standard skip connections pass low-level features directly to the decoder, which can introduce noise and confuse pixel classification.
- Scale Variance: OARs vary significantly in size; models optimized for large structures often fail on tiny ones like the submandibular gland.
- Loss of Global Context: Deep networks often lose the "big picture" spatial relationships necessary for consistent segmentation of complex anatomy.
The SECP-Net Architecture
To address these issues, the authors developed SECP-Net (SE-Connection Pyramid Network). The architecture introduces a multi-stage approach to feature fusion and refinement, combining a primary network for initial estimation and a secondary network for detail enhancement.
SE-Connection (SEC) Module
The SEC module replaces standard concatenation in skip connections. It incorporates Squeeze-and-Excitation (SE) blocks to provide channel-wise attention. This allows the network to adaptively weight important features and suppress background noise, which is particularly beneficial for segmenting small organs with subtle boundaries.
Pyramid Structure and Auto-Context
The network employs a pyramid structure to fuse information across different resolutions. High-semantic features from deeper layers are passed back to shallower layers, ensuring that global context is preserved throughout the decoding process.
Furthermore, SECP-Net uses a Cascaded Auto-context Network. The primary network generates an initial probability map, which is then concatenated with the original image and fed into the secondary network. This "auto-context" approach allows the second stage to focus its attention on the Region of Interest (ROI) identified in the first pass, leading to much finer segmentation boundaries and fewer false positives.
Experimental Evaluation
The model was evaluated using a dataset from the Sun Yat-sen University Cancer Center (SYSUCC), comprising over 40,000 CT images from 356 patients. The dataset includes 13 distinct OAR labels, ranging from the thyroid gland to the brain stem.
Key Results
SECP-Net was compared against several benchmarks, including U-Net, Attention U-Net, U-Net++, CE-Net, and CPF-Net.
- Quantitative Performance: SECP-Net achieved a state-of-the-art average Dice Coefficient of 83.13% and Jaccard Similarity of 76.21%.
- Small Organ Accuracy: Significant improvements were noted for small structures. For instance, the Dice score for the thyroid improved by 2.24% over the baseline U-Net, and the left eye improved by 1.27%.
- Qualitative Refinement: Visual analysis showed that the auto-context mechanism effectively reduced "hallucinated" segmentations, correctly identifying the absence of organs in specific CT slices where other models erroneously predicted their presence.
Conclusion and Future Work
SECP-Net demonstrates that combining channel-wise attention with multi-scale pyramid fusion and cascaded refinement significantly improves OAR segmentation for NPC. While the model handles global context well, future work will focus on integrating Transformer-based backbones (like Swin-Unet) to further enhance spatial relationship modeling and exploring more localized attention mechanisms to better distinguish foreground from cluttered backgrounds in complex CT slices.
Summary updated