Abstract
Precise segmentation of esophageal squamous cell carcinoma (ESCC) is critical for clinical staging and treatment planning. This study presents a comparative analysis of deep learning architectures—UNet, nnUNet, and UMamba—for automated tumor delineation using high-resolution, contrast-enhanced 3D magnetic resonance imaging (MRI). Our findings indicate that UMamba, which incorporates a novel state-space architecture, significantly outperforms traditional convolutional neural networks (CNNs). UMamba achieved a superior Dice Similarity Coefficient (DSC) of 0.764 and reduced boundary errors, highlighting the necessity of capturing global 3D context for accurate esophageal tumor mapping.
Introduction
Esophageal squamous cell carcinoma (ESCC) is an aggressive malignancy where accurate tumor "mapping" or segmentation is vital for determining surgical margins and designing radiation therapy fields. Traditionally, this process is performed manually by radiologists, which is time-consuming and prone to inter-observer variability.
Automating ESCC segmentation is challenging due to the irregular shapes of tumors, heterogeneous internal textures, and proximity to complex anatomical structures in the mediastinum. While computed tomography (CT) has been the standard, this study utilizes contrast-enhanced free-breathing 3D-GRE MRI. This modality offers superior soft-tissue contrast but requires advanced algorithms to effectively process high-resolution volumetric data.
Methods
Dataset and Study Design
The study included a cohort of 192 patients with histologically confirmed ESCC. The dataset was split into 171 patients for model training and 21 patients for independent validation. Imaging was conducted on 3T MRI systems using a contrast-enhanced 3D-GRE sequence with 1 mm isotropic resolution. Ground truth annotations were established through a rigorous process involving two radiologists and a senior expert to resolve discrepancies.
Model Architectures
Four deep learning configurations were evaluated:
- UNet: The standard baseline for medical image segmentation.
- nnUNet (2D & 3D): A self-configuring framework that optimizes hyperparameters automatically.
- UMamba: The primary focus of the study. Unlike standard CNNs that utilize local receptive fields, UMamba integrates a State Space Model (SSM). This allows the model to capture long-range dependencies and global contextual information across the 3D volume more efficiently than Transformers or deep CNNs.
Results
The quantitative evaluation demonstrated a clear advantage for the state-space approach. UMamba achieved the highest accuracy across all primary metrics:
- Dice Similarity Coefficient (DSC): UMamba reached 0.764, outperforming nnUNet-3D (0.738), nnUNet-2D (0.702), and the baseline UNet (0.665).
- Boundary Precision: UMamba significantly reduced the Hausdorff Distance (HD) to 5.048 mm, compared to 5.339 mm for nnUNet-3D and 8.462 mm for the standard UNet. This indicates that UMamba’s predicted boundaries are much closer to the expert manual annotations.
The Importance of 3D Context
A key finding was the performance gap between 2D and 3D architectures. The 3D version of nnUNet showed a 5.1% improvement in DSC over the 2D version. This confirms that capturing the vertical continuity of the esophagus is essential for accurate segmentation, a feature that slice-by-slice 2D models fail to fully utilize.
Discussion and Conclusion
While UMamba provides superior accuracy, it involves a computational trade-off. Its inference time was approximately 5.36 seconds per volume, compared to 0.79 seconds for the 2D UNet. However, this remains significantly faster than manual segmentation and is well within the requirements for clinical workflows.
Limitations and Future Directions
This study was conducted using data from a single institution. Future work must involve multi-center validation to ensure the model's robustness across different MRI protocols and scanner hardware. Additionally, exploring multi-modal models that combine MRI with PET or CT data could further enhance segmentation accuracy by integrating metabolic and anatomical information. Finally, optimizing model efficiency through pruning and quantization will be explored to reduce inference times for 3D architectures.
Summary updated