Reconstructed concatenation in U-shape network based on cross-stage-attention for esophageal segmentation in CT and MR images

Accurate esophagus segmentation in both CT and MR images remains a critical yet unresolved challenge in radiotherapy planning for esophageal carcinoma, primarily due to the organ’s complex tubular morphology, low soft-tissue contrast, and prominent inter-modal variability. Manual delineation is not only time-consuming and labor-intensive but also prone to inter-observer variability, which can compromise dose delivery accuracy and increase the risk of radiation-induced toxicity to adjacent organs at risk (OARs). To address these limitations, this study introduces the Cross-stage-attention Convolution U-Net (CAC-based U-Net), a unified, modality-agnostic neural architecture that reconstructs U-Net’s skip connections using novel Cross-stage-attention Convolution (CAC) blocks. The CAC module synergistically combines large-kernel depth-wise convolutions for capturing global contextual features with small-kernel units for refining local boundary details, effectively replacing naive concatenation and mitigating structural degradation in hierarchical feature maps. Validated on a clinically acquired dataset comprising 29 patients (2383 CT and 3533 MR images), Dense-UNet-CAC and Res-UNet-CAC achieved state-of-the-art segmentation accuracy, with Dice coefficients of 84.02%±8.83% (CT) and 74.08%±14.59% (MR), respectively. The unified framework eliminates the necessity for developing separate modality-specific segmentation pipelines, thereby reducing algorithmic development and maintenance costs while ensuring robust performance across imaging modalities. Its plug-and-play compatibility with backbone networks (e.g., Dense-UNet, Res-UNet) facilitates seamless integration into existing radiotherapy workflows without retraining. External validation on The Cancer Imaging Archive (TCIA) datasets confirmed strong generalizability (Dice coefficients=77.14%±14.59%), underscoring its clinical viability for multi-center deployment. The lightweight design further enhances accessibility in resource-constrained settings, offering a clinically adaptable, cost-effective solution for automated multi-modal esophagus segmentation to improve radiotherapy planning efficiency and anatomical targeting precision.