Abstract
Nasopharyngeal carcinoma (NPC) is a prevalent malignancy in regions such as Southern China and Southeast Asia. Precise segmentation of the target region in MRI scans is critical for effective radiotherapy. We propose the DCTR U-Net, a novel deep learning architecture that integrates Dilated Convolution, Transformer, and Residual modules. This hybrid approach addresses the limitations of standard convolutional neural networks by expanding the receptive field and modeling long-range dependencies. Validated on a clinical dataset of 300 patients, the DCTR U-Net achieved a Dice Similarity Coefficient (DSC) of 0.852 and an Average Symmetric Surface Distance (ASSD) of 0.544 mm, significantly outperforming several state-of-the-art baselines.
Introduction
Nasopharyngeal carcinoma (NPC) is highly sensitive to radiation, making radiotherapy the primary treatment modality. However, the success of radiotherapy is contingent upon the accurate delineation of the tumor boundaries. Currently, this process is largely manual, time-consuming, and prone to inter-observer variability, with misdiagnosis rates in some regions exceeding 40%.
Automated segmentation using deep learning, particularly the U-Net architecture, has shown promise. However, standard U-Nets often struggle to balance local precision (fine boundary details) with global context (spatial relationships with surrounding organs). To overcome these challenges, we developed the DCTR U-Net.
Methods: The DCTR U-Net Architecture
The DCTR U-Net follows an encoder-decoder structure, enhanced by specialized DCTR Blocks. These blocks combine three key paradigms to bridge the gap between local and global feature extraction:
- Dilated Convolution Module: Standard convolutions have a fixed, local receptive field. Dilated (or "atrous") convolutions introduce gaps between kernel elements, allowing the network to capture a broader context without increasing the number of parameters or losing spatial resolution through down-sampling.
- Transformer Module: By incorporating Multi-head Self-Attention (MSA) in the bottleneck and encoder layers, the model can capture long-range dependencies across the entire image, ensuring a global understanding of the anatomy.
- Residual Module: Skip connections are utilized to facilitate the training of deep networks by allowing gradients to flow more easily, effectively learning the residual mapping between layers and preventing the vanishing gradient problem.
Experimental Setup
The model was evaluated using a clinical dataset of 300 NPC patients from Hainan Provincial People’s Hospital (2020–2022). The dataset consisted of 2D MRI slices (512x512 pixels), with ground truth labels provided by experienced radiologists with over five years of experience. A robust 10-fold cross-validation strategy was employed to ensure the reliability and generalizability of the results.
Results and Discussion
The DCTR U-Net was benchmarked against five mainstream models: U-Net, MultiResUNet, TransUNet, Swin-Unet, and UNETR.
- Quantitative Performance: The proposed model achieved a DSC of 0.852, significantly higher than the standard U-Net (0.772) and TransUNet (0.807). The ASSD was reduced to 0.544 mm, indicating superior boundary precision compared to the baseline U-Net's 0.823 mm.
- Qualitative Analysis: Visual inspections confirmed that the DCTR U-Net generates contours that align closely with expert annotations, particularly in complex regions where tumor boundaries are poorly defined or faint.
The synergy between dilated convolutions (for multi-scale feature extraction) and Transformers (for global context) proves essential for medical image segmentation tasks where both fine detail and anatomical context are paramount.
Conclusion
The DCTR U-Net offers a significant advancement in the automated segmentation of NPC. By effectively fusing local and global features, the model provides a more accurate tool for clinical radiotherapy planning. Future work will focus on extending the architecture to 3D volumes and validating the model across multi-center datasets to enhance its generalizability across different patient populations and imaging hardware.
Summary updated