CT-based radiomics for preoperative differentiation of renal cell carcinoma subtypes: internal and external validation
- VernacularTitle:Бөөрний эсийн өмөнгийн хэв шинжүүдийг компьютерт томографид суурилсан радиомик шинжилгээгээр хагалгааны өмнө ялган оношлох асуудалд
- Author:
Enkhbold S
1
;
Purevdorj O
2
;
Telmen K
2
;
Sodgerel B
2
;
Shiirevnyamba A
1
;
Gonchigsuren D
1
;
Oyunbold L
2
Author Information
1. Mongolian National University of Medical Sciences
2. Diagnostic Imaging Center, The First Central Hospital of Mongolia
- Publication Type:Journal Article
- Keywords:
Renal cell carcinoma;
Radiomics;
Computed tomography;
Clear cell;
GLRLM;
Machine learning;
Internal validation;
External validation
- From:
Mongolian Journal of Health Sciences
2026;95(5):37-43
- CountryMongolia
- Language:Mongolian
-
Abstract:
Background:Renal cell carcinoma (RCC) accounts for 85–90% of all primary renal malignancies in adults and 1–3% of all cancers worldwide, with incidence rising steadily year by year. Clear cell (ccRCC), papillary (pRCC), and chromophobe (chRCC) subtypes differ significantly in genetic alterations, biological behavior, and treatment response, making preoperative subtype identification clinically important. Conventional qualitative CT assessment is limited, and contrast-enhanced CT-based radiomics may represent a novel diagnostic approach.
Aim:To develop a model for preoperative differentiation of ccRCC from non-ccRCC subtypes using contrast-enhanced CT-based radiomics, and to evaluate its diagnostic performance through internal and external validation.
Materials and Methods:A retrospective study was conducted on 121 patients with histopathologically confirmed RCC treated at the Urology–Andrology Center of the First Central Hospital of Mongolia between 2021 and 2024. Patients were categorized into ccRCC (n=84) and non-ccRCC (n=37) groups. Tumor segmentation was performed on corticomedullary phase CT images using 3D Slicer software, and a total of 122 radiomic features were extracted, including first-order uniformity, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), shape flatness, and GLCM maximum probability. Diagnostic performance was assessed by ROC analysis, and model generalizability was evaluated through external validation (n=73).
Results:GLRLM demonstrated a statistically significant difference between the ccRCC (mean 0.1230) and non-ccRCC (mean 0.0170) groups (p<0.001). ROC analysis yielded an optimal cut-off value of 0.1493 for GLRLM, with an internal validation AUC of 0.9103, sensitivity of 83.3%, and specificity of 92.3% (p<0.01). External validation (n=73) achieved an AUC of 0.860 and an accuracy of 86.0%, confirming the model’s generalizability. First-order uniformity (AUC=0.880) and variance (AUC=0.862) showed comparable discriminative performance to GLRLM, whereas flatness (AUC=0.584) and GLCM Imc2 (AUC=0.621) demonstrated relatively poor discriminative ability. Confusion matrix analysis yielded an accuracy of 86.0% in both internal and external validation cohorts.
Conclusion:Non-invasive CT-based radiomics demonstrates the ability to accurately differentiate RCC subtypes preoperatively with high precision (AUC=0.92), as confirmed by both internal and external validation.
- Full text:2026092722505264393Бөөрний эсийн өмөнгийн хэв.pdf
