Related Experiment Video
Updated: Jan 11, 2026
In Vivo, Percutaneous, Needle Based, Optical Coherence Tomography of Renal Masses
Published on: March 30, 2015
Tabular prior-data fitted network in real-world CT radiomics: benign vs. malignant renal tumor classification
Tianzhu Liu1, Huanjun Wang2, Yan Guo2
1Department of Radiology, Guangdong Provincial Hospital of Traditional Chinese Medicine, Zhuhai, China.
Background:
Radiomics-based machine learning (ML) models can potentially distinguish between benign and malignant renal tumors on computed tomography (CT). Traditional algorithms necessitate intensive hyperparameter tuning and large datasets for optimization. This study assessed the novel tabular prior-data fitted network (TabPFN), a pre-trained transformer model for tabular data classification, and compared it with conventional algorithms using real-world CT radiomics datasets.
Methods:
Two retrospective cohorts were analyzed: datasets A [207 cystic renal masses (CRMs): 92 benign, 115 malignant] and B [92 tumors: 41 renal oncocytomas (ROs), 51 chromophobe renal cell carcinomas (CRCCs)]. Radiomic features were extracted from three-phase CT images (unenhanced, corticomedullary, and nephrographic) using PyRadiomics. Seven algorithms [support vector machine (SVM), stochastic gradient descent (SGD), k-nearest neighbor (KNN), random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and TabPFN] were evaluated via 10-fold cross-validation. Evaluation metrics included area under the curve (AUC), accuracy, sensitivity, specificity, and feature importance.
Results:
In dataset A, all models achieved high AUCs (training: 0.935-1.000; validation: 0.800-0.946). TabPFN showed top-tier performance, particularly in nephrographic-phase analysis (validation AUC: 0.946). In dataset B, TabPFN demonstrated superior stability (validation AUC: 0.700-0.800), outperforming SVM and KNN while circumventing the convergence failures of SGD. On feature importance analysis, its dynamic weight allocation paralleled that of ensemble models (RF/XGBoost) without explicit hyperparameter optimization.
Conclusions:
TabPFN demonstrates robust performance in differentiating renal tumors, particularly in small-scale, high-dimensional datasets. Its transformer-based architecture and pre-training on synthetic data eliminate manual parameter tuning, enhancing clinical applicability over conventional ML approaches.
