Pilot Study on Contrast-Enhanced CT-Based 3D Cascaded Segmentation of Hypopharyngeal Cancer and the Stability of the Extracted Radiomics Features.
Authors
Affiliations (7)
Affiliations (7)
- College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, 110819, Liaoning, China.
- Department of Otolaryngology, The First Affiliated Hospital of China Medical University, Shenyang, 110001, Liaoning, China.
- Department of Chinese Medicine, The First Affiliated Hospital of China Medical University, Shenyang, 110001, Liaoning, China.
- Department of Otolaryngology, The First Affiliated Hospital of China Medical University, Shenyang, 110001, Liaoning, China. [email protected].
- College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, 110819, Liaoning, China. [email protected].
- Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, Shenyang, 110819, Liaoning, China. [email protected].
- Department of Chinese Medicine, The First Affiliated Hospital of China Medical University, Shenyang, 110001, Liaoning, China. [email protected].
Abstract
Hypopharyngeal cancer (HPC), a malignant head and neck tumor with poor prognosis, requires accurate pretreatment segmentation for surgical planning and radiomics analysis. However, no prior research has reported automated segmentation of hypopharyngeal cancer on CECT images or investigated the impact of such segmentation on radiomics features. This study developed a 3D cascaded HPCNet framework for automated HPC segmentation on CECT images and evaluated the stability of extracted radiomics features. Pretreatment CECT images of 180 patients with HPC were retrospectively acquired, and a 5-fold cross-validation strategy was adopted for dataset division. The cascaded HPCNet framework significantly outperformed the nnUNet in the segmentation of HPC, achieving the highest median DSC (0.793 vs. 0.768, p <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 0.05), Jaccard index (0.650 vs. 0.635, p <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 0.05), precision (0.808 vs. 0.797, p <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 0.05), and recall (0.795 vs. 0.778, p <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 0.05), particularly when the tumor volume was <math xmlns="http://www.w3.org/1998/Math/MathML"><mo>≤</mo></math> 10 <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>c</mi> <msup><mi>m</mi> <mn>3</mn></msup> </mrow> </math> (DSC: 0.796 vs. 0.762, p <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 0.05). The maximum tumor diameter predicted by the cascaded HPCNet framework was highly correlated with the maximum tumor diameter delineated manually (p <math xmlns="http://www.w3.org/1998/Math/MathML"><mo><</mo></math> 0.0001). Five of the seven extracted radiomics feature groups exhibited high correlation, with ICC <math xmlns="http://www.w3.org/1998/Math/MathML"><mo>></mo></math> 0.80. In conclusion, the proposed cascaded HPCNet framework can achieve high-precision automatic segmentation of HPC on CECT images, and the extracted radiomics features exhibit a high level of stability.