Radiomics models based on enhanced computed tomography to distinguish clear cell from non-clear cell renal cell carcinomas.

Sci Rep 2021 AI 7 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Why Subtype Matters in Kidney Cancer

Renal cell carcinoma (RCC) is the predominant kidney cancer, accounting for 2 to 3 percent of adult malignant tumors worldwide. Its main subtypes are clear cell RCC (ccRCC), which comprises roughly 70 percent of cases, and the less common non-clear cell forms including papillary, chromophobe, and collecting duct carcinoma.

The distinction between ccRCC and non-ccRCC is clinically critical because the two groups respond very differently to targeted therapies. ccRCC carries a higher rate of metastasis, and its prognosis without treatment is considerably worse than non-ccRCC. Early and accurate subtype identification directly shapes which drugs a patient receives.

Enhanced CT scanning is the primary imaging tool for diagnosing RCC, but even experienced radiologists misclassify subtypes at a meaningful rate. No blood-based tumor markers can reliably separate ccRCC from non-ccRCC, leaving a diagnostic gap that this study aimed to address.

Renal biopsy is the definitive standard but introduces real risks including tumor seeding, hemorrhage, and sampling error. A non-invasive imaging-based approach that could reliably classify RCC subtypes would reduce the need for biopsy and enable faster clinical decision-making.

TL;DR: Accurately distinguishing ccRCC from non-ccRCC at diagnosis matters because the two subtypes require different targeted therapies, yet current CT interpretation misclassifies cases at a significant rate.
Pages 1-2
What Radiomics Adds to Imaging

Radiomics is a field that extracts quantitative features from medical images that cannot be perceived by the human eye. Since the concept was formally defined in 2012, researchers have shown that these hidden features correlate with tumor biology, metastatic potential, and patient survival across multiple cancer types.

The features extracted through radiomics include measures of texture, shape, intensity distribution, and higher-order statistical patterns. When computed from CT images, these values capture information about tumor heterogeneity, vascular supply, and microenvironment that conventional visual review misses.

Previous studies applied radiomics to distinguish benign from malignant renal masses, but few had attempted to differentiate ccRCC from non-ccRCC using this approach. This study was designed to fill that gap by building and comparing three machine learning classifiers on radiomics features extracted from cortical-phase CT images.

TL;DR: Radiomics converts CT scan pixel data into hundreds of quantitative features encoding tumor texture and heterogeneity, giving machine learning models information that trained radiologists cannot see with the naked eye.
Pages 2, 5, 6
Study Design and Feature Extraction

The study enrolled 190 patients with pathologically confirmed RCC, including 147 ccRCC and 43 non-ccRCC cases. Patients were randomly split 70/30 into training (132 patients) and testing (58 patients) sets, with demographic and subtype distributions balanced across both groups.

CT images were acquired on a 64-row scanner with renal cortical, medullary, and excretion phases. Radiologists with 15 to 20 years of experience manually outlined the tumor volume of interest layer by layer on cortical-phase images. Inter- and intra-observer reproducibility was measured using intraclass correlation coefficients, yielding a mean ICC of 0.78, indicating good agreement.

From each patient's outlined tumor region, 396 radiomic features were extracted covering histogram statistics, Haralick texture measures, shape factors, gray-level size zone matrix, gray-level co-occurrence matrix, and run-length matrix features. All features were standardized before selection.

Feature selection used ANOVA, general linear modeling, and mutual information filtering to reduce from 396 features to a shortlist, then correlation analysis and logistic regression culled the list to four stable, non-redundant features: Variance, HighGreyLevelRunEmphasis, MinIntensity, and OneVoxelVolume. These four served as inputs to all three machine learning models.

TL;DR: Starting from 396 features extracted by 3D texture analysis of cortical-phase CT images, a rigorous selection pipeline identified four stable radiomic features that were fed into three machine learning classifiers.
Pages 3-4
Model Performance Versus Radiologist Diagnosis

Three classifiers were trained and evaluated: Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR). On the held-out testing set, the LR model achieved an AUC of 0.906, sensitivity of 0.956, and specificity of 0.692. The RF model reached AUC 0.909 with sensitivity 0.956, while SVM achieved AUC 0.841 with perfect sensitivity but very low specificity.

In direct comparison, trained radiologists reviewing the same contrast-enhanced CT images achieved a sensitivity of 0.850 and specificity of 0.581 with an AUC of only 0.69. All three radiomic models outperformed radiologist-level diagnosis on the key metric of AUC.

The LR model was identified as the overall best performer due to its combination of high AUC, highest specificity among the three models, and a strong F1 score of 0.935 and positive predictive value of 0.915. Specificity is particularly important in this task, where false positives (wrongly calling ccRCC when it is not) can misdirect treatment.

The radiomic results also surpassed a quantitative multiparametric MRI study (AUC 0.804) and outperformed the tumor biomarker GPX1 (AUC 0.79) on the same classification task. This supports radiomics as a competitive non-invasive diagnostic approach for RCC subtyping.

TL;DR: All three radiomic classifiers exceeded radiologist-level CT interpretation for distinguishing ccRCC from non-ccRCC, with the logistic regression model reaching an AUC of 0.906 compared to the radiologist AUC of 0.69.
Pages 4-5
Tumor Heterogeneity and Why 3D Analysis Matters

Tumor heterogeneity is a hallmark of malignancy that influences growth behavior, metastatic risk, and response to therapy. Heterogeneity in tumor vasculature can drive hypoxia, increased interstitial pressure, and greater invasiveness. Radiomic texture features capture this heterogeneity in ways that visual inspection and simple size measurements cannot.

Most prior radiomics studies in RCC relied on two-dimensional texture analysis of selected CT slices. This study adopted a 3D volume of interest (VOI) approach, delineating the entire tumor volume layer by layer. Three-dimensional analysis is more comprehensive and reduces the risk of missing heterogeneous regions present only in some tumor planes.

The four selected features reflect meaningful biological variation. Variance and gray-level run-length measures capture differences in pixel intensity distribution and repetition patterns that correlate with vascular density and tissue composition differences between ccRCC and non-ccRCC. MinIntensity and OneVoxelVolume reflect baseline enhancement and spatial texture properties.

TL;DR: Three-dimensional radiomic feature extraction across the full tumor volume captures heterogeneity information that is missed by two-dimensional slice analysis, providing a more biologically meaningful input to machine learning classifiers.
Pages 4-5
Limitations and What They Mean for Future Studies

The study population of 190 patients, drawn exclusively from a single Chinese center, is a recognized limitation. The relatively small number of non-ccRCC cases (43 total) created a class imbalance that contributed to lower specificity across all models. Imbalanced datasets make it harder for classifiers to distinguish the minority class reliably.

Techniques such as SMOTE (Synthetic Minority Oversampling TEchnique) have been proposed to address class imbalance, but these methods can introduce artificial data points that do not reflect real tumor biology. The authors chose not to use them, prioritizing the integrity of the training data.

The retrospective single-center design limits generalizability, and the study focused only on Chinese patients. Prospective, multicenter trials using larger and more ethnically diverse cohorts, combined with novel imaging markers, would be needed before radiomic subtyping could enter routine clinical use.

TL;DR: The small sample size and single-center, retrospective design limit immediate clinical application, but the results justify prospective multicenter radiomic trials for RCC subtyping.
Page 5
Radiomics as a Non-Invasive Diagnostic Tool

This study demonstrated that CT-based radiomic models can meaningfully improve the non-invasive differentiation of ccRCC from non-ccRCC. By extracting 3D texture features and training machine learning classifiers, the approach exceeded both human radiologist performance and previously reported MRI and biomarker-based methods.

The practical implication is that radiomics could serve as an adjunct to standard CT interpretation, flagging cases likely to be non-ccRCC before treatment decisions are made. This would allow appropriate targeted therapy to be chosen without waiting for biopsy results or incurring the risks of invasive tissue sampling.

Future directions include prospective validation, incorporation of multi-phase CT data, and integration with genomic or clinical variables to build combined diagnostic and prognostic models that support precision medicine for kidney cancer patients.

TL;DR: CT radiomics with machine learning outperforms standard radiologist interpretation for RCC subtyping and could serve as a non-invasive adjunct to guide treatment selection in kidney cancer.
Citation: Open Access, 2021. Available at: PMC8253856.