Deep Learning-Based CT Splenic Segmentation and Morphometrics: A Spleen Volume Prediction Model
| Author | Tian Tan | en |
| Author | Xiaoying Wang | en |
| Author | Yaofeng Zhang | en |
| Author | Xiaodong Zhang | en |
| Author | Kexin Wang | en |
| Orcid | Xiaoying Wang [0000-0001-9822-961X] | en |
| Issued Date | 2025-10-31 | en |
| Abstract | Background: Quantitative assessment of spleen volume on computed tomography (CT) is clinically relevant but is often approximated using linear measurements. Deep learning–based segmentation enables automated volumetry; however, population-specific normative spleen volume models for Chinese adults remain limited. Objectives: To develop a deep learning–based automated CT segmentation model for normal spleens and to derive a Chinese adult standard spleen volume (SSV) prediction model based on key physiologic factors. Patients and Methods: A 3D V-Net spleen segmentation model was trained using Dataset 1 (training, n = 3418; validation, n = 413; test, n = 443). Internal validation used Dataset 2 from our institution (n = 1996; January-April 2024), and external validation used 2809 publicly available thin-slice CT examinations. Model performance was evaluated using the Dice similarity coefficient (DSC), Hausdorff distance (HD), and volume similarity (VS). For normative modeling, spleen volumes from 578 healthy adults on portal venous phase CT were analyzed, and candidate prediction variables were assessed using correlation and regression analyses. Results: Segmentation performance was high on the test set {DSC, 0.988 (0.984 - 0.989) [median (IQR)]; VS, 0.997 (0.994 - 0.999); HD, 0.017 (0.014 - 0.021)} and remained robust on external validation [DSC, 0.982 (0.974 - 0.987)]. Volume agreement analyses showed a mean absolute error of 1.945 mL (test set), 3.829 mL (internal validation), and 5.806 mL (external validation). In the cohort of 578 adults, the spleen volumes ranged from 51.09 to 644.38 cm³, with a median of 177.90 cm³. The median (IQR) of the x-, y-, and z-axis diameters were 8.71 (7.97 - 9.53) cm, 9.20 (8.20 - 10.46( cm, and 9.10 )8.00 - 10.30( cm, respectively. Age (r = -0.24, P < 0.0001) and gender (male = 0, female = 1; r = -0.32, P < 0.0001) were negatively correlated with splenic volume (SV), while height (r = 0.35, P < 0.0001), weight (W; r = 0.45, P < 0.0001), Body Mass Index (BMI; r = 0.32, P < 0.0001), and body surface area (BSA; r = 0.46, P < 0.0001) were positively correlated with SV. Based on thin-slice portal venous phase CT images, a bidirectional stepwise selection procedure identified body surface area (BSA) and age as significant predictors. The final model was: log (SSV) = 3.708 + 0.987 × BSA − 0.00629 × Age (R² = 0.269). Conclusion: A 3D V-Net model enabled accurate automated spleen segmentation with consistent performance across internal and external validation cohorts. An SSV prediction model based on BSA and age provides a population-specific reference for quantitative spleen volumetry in Chinese adults. | en |
| DOI | https://doi.org/10.5812/iranjradiol-168436 | en |
| URI | https://brieflands.com/journals/ijradiology/articles/168436 | en |
| Keyword | Standard Splenic Volume (SSV) | en |
| Keyword | Morphometric | en |
| Keyword | Deep Learning | en |
| Keyword | Computed Tomography | en |
| Publisher | Brieflands | en |
| Title | Deep Learning-Based CT Splenic Segmentation and Morphometrics: A Spleen Volume Prediction Model | en |
| Type | Research Article | en |