Angiomyolipoma (AML) is the most common benign solid renal neoplasm, while renal cell carcinoma (RCC) accounts for nearly 90% of renal malignancies. Their clinical management is fundamentally different: RCC typically requires partial or radical nephrectomy, while AML is generally observed or embolized. Correctly distinguishing the two on imaging is therefore critical to avoid unnecessary surgery.
On conventional B-mode ultrasound, AML is classically hyperechoic, but minimal-fat AML may appear isoechoic, and 32% of RCCs under 3 cm are hyperechoic, directly mimicking AML. Grayscale ultrasound achieves only 42% diagnostic accuracy for small solid renal lesions, leaving many patients requiring CT or MRI follow-up to resolve diagnostic uncertainty.
Point shear wave elastography (pSWE) uses acoustic radiation force impulse technology to quantitatively measure tissue stiffness through shear wave velocity (SWV), offering an operator-independent, non-invasive addition to standard ultrasound. However, prior pSWE studies for RCC versus AML differentiation showed highly conflicting results, ranging from 88% sensitivity and 54% specificity to only 48% sensitivity and 33% specificity.
Machine learning has been successfully applied to elastography in other organs, but had never been evaluated for solid renal lesion characterization using pSWE. This study hypothesized that machine learning applied to composite statistical features from multiple tissue regions would substantially outperform simple median SWV thresholding for RCC versus AML classification.
This prospective, IRB-approved, single-center study enrolled 51 patients with 52 renal tumors (42 RCCs confirmed by surgical pathology, 10 AMLs confirmed by pathology, CT, or MRI) between February 2014 and May 2016. All examinations were performed on a Siemens Acuson S2000 system by certified sonographers with at least 18 months of clinical elastography experience.
Ten consecutive valid SWV measurements were obtained from three anatomically distinct regions for each patient: the renal tumor, the ipsilateral renal cortex, and the ipsilateral renal medulla. From each location's ten measurements, four statistical features were derived -- mean, median, interquartile range, and standard deviation -- yielding up to twelve features total when all three regions were combined.
Four supervised machine learning algorithms were evaluated: logistic regression, naive Bayes, quadratic discriminant analysis (QDA), and nonlinear support vector machines (SVM) with a Gaussian radial basis function kernel. Each algorithm was trained and tested using leave-one-out cross-validation, and model performance was reported as area under the ROC curve (AUC) with Wilcoxon rank-sum p-values for class separation significance.
Models were evaluated across four input configurations: tumor features only, cortex features only, medulla features only, and the full 12-feature combination of all three regions. The DeLong method was used to statistically compare AUC values between models, and baseline performance was established using median tumor SWV and tumor-to-cortex and tumor-to-medulla shear wave ratios.
Median tumor SWV performed poorly as a standalone classifier, achieving AUC of only 0.62 (p = 0.23), confirming that a single summary statistic from the tumor alone is insufficient for reliable RCC versus AML differentiation. Tumor-to-cortex shear wave ratio achieved AUC of 0.64, and tumor-to-medulla ratio achieved AUC of 0.72, both below clinically useful thresholds.
Support vector machines outperformed all other algorithms and achieved statistically significant separation of RCC from AML using features from any single region: tumor alone (AUC = 0.94, p = 4.6e-3), cortex alone (AUC = 0.79, p = 2.3e-5), or medulla alone (AUC = 0.84, p = 1.1e-3). No other algorithm reached statistical significance with single-region inputs.
Combining all 12 statistical features from tumor, cortex, and medulla yielded the highest SVM performance at AUC = 0.98 with 94% accuracy (p = 3.1e-6), representing a highly significant improvement over median SWV (p = 2.8e-4 for the difference). QDA also reached statistical significance with combined features (AUC = 0.93), while naive Bayes achieved AUC = 0.78. Logistic regression was the only algorithm that failed to reach significance even with all 12 features (AUC = 0.71, p = 0.099).
The distribution analysis of SWV measurements revealed that AML measurements cluster more tightly around the median, while RCC measurements show greater spread, explaining why incorporating variance-based features (standard deviation, interquartile range) in addition to median substantially improved machine learning classification relative to median-only analysis.
The superior performance of SVM over other algorithms is consistent with the known strength of nonlinear kernels for high-dimensional feature spaces. The Gaussian radial basis function kernel enables SVM to identify complex, non-linear decision boundaries between RCC and AML that linear classifiers cannot capture from the 12-dimensional feature space.
The diagnostic value of SWV measurements outside the tumor -- in the renal cortex and medulla -- is biologically plausible. Malignant tumors alter the mechanical properties of surrounding parenchyma through local invasion, edema, and perfusion changes. Prior animal studies have demonstrated that experimentally altering renal perfusion through arterial or venous clamping changes SWV values, supporting the hypothesis that RCC-induced vascular changes in adjacent tissue alter local stiffness.
This study is the first to apply machine learning to solid renal lesion characterization using pSWE, and it differs from prior elastography machine learning studies by using composite statistical features from different anatomical regions rather than directly analyzing stiffness color maps. This approach captures tissue heterogeneity information across regions that color-map-based methods may miss.
The conflicting prior results in pSWE for renal lesions are explained in part by the use of single-metric approaches: median tumor SWV captures only central tendency and discards the distributional information encoded in variance and percentile features that the machine learning model exploits.
This study demonstrates that machine learning applied to multi-region pSWE features can achieve near-diagnostic-quality classification of RCC versus AML, with SVM accuracy of 94% and AUC of 0.98 using prospectively collected data from 51 patients.
The finding that measurements from the renal cortex and medulla contribute significantly to classification performance establishes that pSWE should be obtained from multiple anatomical regions during routine renal ultrasound examinations to maximize its diagnostic utility, rather than restricting measurements to the lesion itself.
Future studies should enroll larger patient cohorts, include additional tumor types such as oncocytoma, chromophobe RCC, and papillary RCC, and assess inter-observer variability across multiple operators to validate the clinical robustness of this approach before routine clinical deployment.
If validated in multi-institutional studies, pSWE with machine learning could be integrated into standard renal ultrasound protocols to provide real-time, non-invasive classification of incidentally detected renal masses, potentially reducing unnecessary CT, MRI, and surgical procedures for patients with benign AML.