TL;DR
This study leverages CT radiomics to distinguish between Wilms tumor (WT) and clear cell sarcoma of the kidney (CCSK) in pediatric patients, achieving promising diagnostic performance. By analyzing quantitative imaging features beyond human visual perception, this approach offers a non-invasive tool to aid in the differential diagnosis of these rare but critical pediatric renal malignancies.
What Problem Does This Paper Address?
Wilms tumor (WT) and clear cell sarcoma of the kidney (CCSK) are the two most common types of pediatric renal malignancies. While WT is more prevalent, CCSK often presents with a less favorable prognosis due to its tendency for bone metastases and late recurrence. Differentiating between these two entities is challenging, as their clinical and radiological presentations can be similar. Current diagnostic methods often rely on histopathological analysis and immunohistotyping, which have limitations. Biopsies only sample a small portion of the tumor, and the diverse histology of CCSK can complicate diagnosis. This study aims to address this diagnostic challenge by exploring the utility of CT radiomics, a technique that extracts a large number of quantitative features from medical images, to differentiate between WT and CCSK.
Methods
The researchers employed a retrospective approach, analyzing contrast-enhanced CT scans from pediatric patients diagnosed with either WT or CCSK. The core of their methodology involved a radiomics workflow:
- Image Segmentation: Tumors were manually delineated on corticomedullary phase CT images by a radiologist, with a senior radiologist reviewing and verifying these segmentations. To ensure reproducibility, a subset of cases underwent a second segmentation, and intra-class correlation coefficients (ICCs) were calculated.
- Feature Extraction: Radiomics features were extracted from the segmented tumors using the FeAture Explorer software. These features were categorized into three groups: first-order (voxel intensity distribution), shape, and texture (including GLCM, GLRLM, GLSZM, NGTDM, and GLDM). A total of 107 radiomics features were initially extracted per lesion.
- Dimensionality Reduction and Feature Selection: To manage the high dimensionality of radiomics features and improve model robustness, several steps were taken:
- Normalization: Features were normalized using z-score standardization.
- Reproducibility Filtering: Features with an ICC below 0.80 between double annotations were removed.
- Correlation Filtering: Features with a Pearson correlation coefficient (PCC) exceeding 0.90 were eliminated to reduce redundancy.
- LASSO Regression: The Least Absolute Shrinkage and Selection Operator (LASSO) method was used for feature selection, identifying a subset of important features.
- Multivariate Stepwise Logistic Regression: This method was employed to select the final optimal radiomics features with statistical significance (P < 0.05).
- Model Building: A logistic regression model was constructed using the selected radiomics features. The model's performance was evaluated using metrics such as the Area Under the Curve (AUC), accuracy, sensitivity, specificity, and F1 score. Five-fold cross-validation was used for internal validation within the training set, and the model was then tested on an independent test set.
- Image Analysis: In addition to radiomics, qualitative and semi-quantitative imaging features were also analyzed by an experienced pediatric radiologist. These included maximum tumor diameter, the ratio of maximum CT value of the tumor solid portion to the mean CT value of the contralateral renal vein (CTmax/CT renal vein), and the presence of dilated peritumoral cysts.
Datasets
The study utilized a dataset of 116 pediatric patients with confirmed diagnoses of either Wilms tumor (WT) or clear cell sarcoma of the kidney (CCSK). These patients were retrospectively collected from March 2011 to October 2022. The dataset was randomly stratified into a training set (n=81) and a test set (n=35) using a 7:3 ratio. The training set comprised 58 WT cases and 23 CCSK cases, while the test set included 25 WT cases and 10 CCSK cases. No new benchmark datasets were introduced.
Key Insights and Ideas
- Radiomics Captures Subtle Differences: The study highlights that quantitative radiomics features are more adept at capturing subtle imaging differences between CCSK and WT than visual analysis alone. This suggests that radiomics can unlock information hidden within CT scans that is not readily apparent to the human eye.
- Beyond Traditional Imaging: While traditional qualitative and semi-quantitative imaging features (like tumor diameter and CTmax/CT renal vein ratio) showed some differences in the training set, these were not consistently significant in the test set. This underscores the need for more advanced analytical methods like radiomics.
- Robust Model Performance: The developed radiomics model demonstrated strong and consistent performance across both the training and test sets, indicating its potential for generalizability. The model's ability to maintain favorable diagnostic metrics even with a limited test set is a significant finding.
- Clinical Utility: The study's inclusion of clinical decision curves and clinical impact curves suggests that the radiomics model has practical clinical value, potentially aiding in treatment decisions and reducing diagnostic uncertainty.
Key Results
The radiomics model achieved the following performance metrics:
- Training Set:
- AUC: 0.889 (95% CI: 0.811–0.967)
- Accuracy: 0.864
- Sensitivity: 0.739
- Specificity: 0.914
- Five-fold Cross-validation (Training Set):
- AUC: 0.863 (95% CI: 0.774–0.952)
- Accuracy: 0.852
- Test Set:
- AUC: 0.792 (95% CI: 0.616–0.968)
- Accuracy: 0.857
- Sensitivity: 0.600
- Specificity: 0.960
The study also found statistically significant differences in maximum tumor diameter (P=0.021) and the presence of dilated peritumoral cysts (P=0.005) between WT and CCSK in the training set, though these were not significant in the test set.
Limitations and Future Work
The authors acknowledge several limitations:
- Rarity of Diseases: Both CCSK and WT are rare, with CCSK being particularly uncommon. The study's sample size (116 patients) may raise concerns about statistical power and generalizability.
- Data Augmentation Bias: The use of SMOTE for data augmentation, while helpful for imbalanced datasets, can introduce biases. The authors suggest assessing model performance on independent datasets to gauge these potential biases.
- Lack of External Validation: Due to the rarity and challenges in case collection, external validation with multi-center data was not performed. This is identified as a crucial next step to confirm the model's generalization capability.
Future work should focus on collecting multi-center data for external validation and further exploring the clinical implications of this radiomics approach in the management of pediatric renal tumors.
Summary updated