EsoNet: A Simple and Robust Network for Esophageal Tumor Segmentation with NBI Endoscopic Images

Accurate segmentation of esophageal tumors is essential for computer-aided diagnosis and treatment planning. However, both elevated and flat tumor types pose challenges due to their visual similarity to normal tissue and variability in appearance. We propose EsoNet, a simple yet robust UNetlike architecture tailored for esophageal tumor segmentation in NBI endoscopic images. EsoNet incorporates Robust Feature Extraction (RFE) blocks and their residual variants to enhance feature learning. Evaluations on a private NBI dataset from Kyushu University Hospital show that EsoNet outperforms several U-Net variants, especially in segmenting flat tumors. Ablation studies reveal that group normalization significantly improves segmentation stability and accuracy under limited data conditions. Our method achieves state-of-the-art performance while maintaining architectural simplicity. Code: https://github.com/code4works/esonet