Endometrial cancer is one of the most common gynecological malignancies worldwide, and accurate delineation of tumor boundaries on CT scans is essential for treatment planning, radiotherapy targeting, and clinical staging. Manual segmentation performed by radiologists is time-consuming, subject to inter-observer variability, and difficult to scale across large patient populations.
Automated deep learning segmentation offers the potential to produce consistent, reproducible tumor boundaries rapidly. However, CT images of the uterine region present unique challenges: the target structure has highly irregular shapes, low contrast with surrounding soft tissue, and variable appearance across patients.
Prior deep learning approaches such as standard U-Net architectures have shown promise but often struggle with edge definition and fine structural details in abdominal CT scans. This study introduces SCWU-Net, a new architecture designed specifically to address these shortcomings for endometrial cancer segmentation.
SCWU-Net is built on the standard encoder-decoder U-Net framework but introduces two major innovations: the Spatial Selection Module (SSM) and the Combination Weight Module (CWM). These components work together to capture multi-scale spatial features and enhance edge detection in the skip connections that bridge encoder and decoder layers.
The SSM uses large-kernel convolutions - specifically 5x5 and 7x7 kernels - to capture broader spatial context than the typical 3x3 convolutions used in standard networks. By processing feature maps at different receptive field sizes simultaneously, the module can identify both fine local details and coarser structural patterns within the same layer.
The CWM is applied within the skip connections between encoder and decoder paths. It functions as an attention mechanism that assigns different importance weights to different spatial positions and channels, allowing the network to focus on edge regions and boundary features that are most relevant for precise segmentation. This is particularly important for distinguishing tumor margins from adjacent normal tissue.
The model was evaluated on the ECPC-IDS dataset, a publicly available collection of CT images from endometrial cancer patients. This dataset provides manually annotated ground-truth tumor masks, enabling objective comparison of segmentation accuracy across different methods.
Training followed standard deep learning practices with data augmentation to prevent overfitting on the relatively limited medical imaging dataset. The model was compared against several established architectures including the original U-Net, TransUNet, and Swin-Unet to benchmark performance gains from the proposed modifications.
Evaluation metrics included the Dice Similarity Coefficient (DSC), which measures the overlap between predicted and ground-truth masks, as well as Intersection over Union (IoU), precision, and recall. These metrics collectively capture different aspects of segmentation quality, from overall boundary agreement to sensitivity in detecting true tumor voxels.
SCWU-Net achieved a Dice Similarity Coefficient of 82.98% and an IoU of 78.63% on the ECPC-IDS test set, outperforming all baseline architectures tested. Precision reached 92.36% and recall 84.10%, indicating that the model is both highly accurate in its positive predictions and sensitive in detecting true tumor regions.
These results represent meaningful improvements over standard U-Net and transformer-based alternatives. The gains are particularly notable in boundary regions where tumor margins are irregular or where the tumor abuts other soft tissue structures - exactly the situations where existing methods tend to produce the most errors.
Ablation studies confirmed that both the SSM and CWM modules contribute independently to performance. Removing either module led to measurable drops in DSC and IoU, validating the design choices and demonstrating that the improvements are not attributable to increased parameter count alone.
The success of the SSM's large-kernel convolutions (5x5 and 7x7) relates to a fundamental challenge in medical image segmentation: tumors occupy a wide range of spatial scales within a single image. A 3x3 kernel, while computationally efficient, has a limited receptive field and cannot easily capture the broader structural context that distinguishes tumor tissue from surrounding organs.
Large kernels effectively act as a form of built-in spatial attention, allowing the network to weight distant pixel relationships within a single convolutional operation. This is functionally similar to what transformer-based architectures achieve through self-attention, but with lower computational overhead and better translation of features learned during training to new test images.
The combination of multi-scale kernels within SSM means the network simultaneously processes local edge information and global shape context, producing richer feature representations at each encoder layer that downstream decoder stages can leverage for precise boundary reconstruction.
The strong segmentation performance demonstrated by SCWU-Net positions it as a candidate for integration into clinical radiotherapy planning workflows. Automated, consistent tumor delineation could reduce the workload on radiologists and decrease inter-clinician variability in treatment target volumes, potentially improving the consistency of patient outcomes.
Future work could explore applying SCWU-Net to MRI-based segmentation, where soft tissue contrast is higher than in CT and the architectural principles may show even greater benefit. Extension to multi-organ segmentation - simultaneously delineating the uterus, cervix, and surrounding critical structures - would further increase clinical utility.
Validation on independent multicenter datasets will be essential before clinical deployment, as performance on a single benchmark dataset does not guarantee generalization across different CT scanners, imaging protocols, and patient populations. Nevertheless, this work establishes a strong foundation for AI-assisted endometrial cancer imaging analysis.