Introduction
Automatic segmentation of the cancerous esophagus in computed tomography (CT) images is a critical computer-assisted method for improving the efficiency of diagnosis and treatment planning, particularly in radiotherapy. However, this task is notoriously difficult due to the high anatomical variability of the esophagus across different cancer stages and locations. Furthermore, the low contrast between the esophagus and surrounding soft tissues often results in "blurry" boundaries that are nearly indistinguishable to standard segmentation algorithms.
Existing approaches typically fall into two categories: 2D networks that ignore vital depth (z-axis) information, and 3D networks that, while spatially aware, often suffer from high computational costs and overfitting on limited medical datasets. To bridge this gap, we present Eso-Net, a 2.5D segmentation network designed to leverage 3D structural priors while maintaining the efficiency of 2D processing.
The Eso-Net Architecture
Eso-Net is built on an encoder-decoder architecture and utilizes a 2.5D approach, taking three consecutive CT slices as input to segment the middle slice. This configuration captures local volumetric context without the memory overhead of full 3D models.
Multi-Structure Response Filter (MSRF)
A key innovation in Eso-Net is the Multi-Structure Response Filter (MSRF). Instead of attempting to enhance the irregular cancerous tissue directly, the MSRF uses Hessian-based 3D filters to identify and enhance regular surrounding structures, such as tube-like blood vessels or plate-like membranes. By defining the "normal" geometry of the background, the network can more effectively isolate the "abnormal" and irregular cancerous regions.
Prior Attention Module (PAM)
To integrate these geometric priors, we developed the Prior Attention Module (PAM). Unlike standard self-attention mechanisms that learn spatial relevance from scratch, PAM uses the MSRF output as an explicit spatial prior. This guides the network's attention toward relevant regions during feature transmission through skip connections, significantly improving boundary precision.
Experimental Results
The model was evaluated on a dataset of 30 esophageal cancer patients (6,362 slices) from the Sun Yat-sen University Cancer Center. Using five-fold cross-validation, Eso-Net achieved a Dice Similarity Coefficient (DSC) of 84.839% and a Hausdorff Distance (HD) of 2.583 mm.
As shown in the figures, Eso-Net demonstrates a superior ability to distinguish the esophagus from neighboring tissues that share similar intensities. Quantitatively, it outperformed established benchmarks including U-Net, 3D U-Net, and Attention U-Net across all key metrics.
Conclusion
Eso-Net addresses the primary challenges of esophageal cancer segmentation by combining 2.5D efficiency with 3D structural priors. The "inverse" enhancement strategy of the MSRF provides a robust way to handle irregular tumor shapes. Future work will focus on making the MSRF parameters learnable to improve generalization across different imaging modalities and organ types.
Summary updated