Automated segmentation by SCA-UNet can be directly used for radiomics diagnosis of thymic epithelial tumors.

BACKGROUND

Automatic segmentation of thymic lesions in preoperative computed tomography (CT) images is crucial for accurate diagnosis but remains time-consuming. Although UNet is widely used in medical imaging, its performance is limited by the inherent drawbacks of convolutional neural networks (CNNs), such as restricted receptive fields and limited global context modeling, which affect segmentation efficiency.

METHOD: 712 patients with mediastinal lesions admitted to Shanghai General Hospital between October 2014 and January 2023 were included in the study. Each lesion was manually delineated on CT images using the 3D slicer workstation. To enhance global context awareness for lesion segmentation, previously collected training data was used to develop a deep learning network called Space Channel Attention UNet (SCA-UNet). The model was further utilized for radiomics-based identification and risk assessment of thymic epithelial tumors (TETs). The code of SCA-UNet is available at: https://github.com/GovernTony/SCA-UNet.

RESULT: The SCA-UNet model was developed using 107 selected radiomic features. Based on our CT dataset, SCA-UNet outperformed several state-of-the-art models in segmentation accuracy and generalization, achieving the highest Dice Similarity Coefficient (DSC) of 87.48%. Furthermore, in subsequent radiomics classification, the segmentation results produced by SCA-UNet were comparable to those obtained through manual segmentation, with an Area Under the Curve (AUC) of 0.8457 for SCA-UNet versus 0.8514 for manual segmentation in TET identification, and 0.7735 for SCA-UNet versus 0.7780 for manual segmentation in risk assessment.

CONCLUSION

Overall, SCA-UNet demonstrated high accuracy in automatic segmentation and can be effectively applied to radiomics analysis, showing significant potential for the clinical application of TET treatment.