用于区分侵袭性与惰性肾肿瘤的MRI影像组学特征

Renal tumors are highly heterogeneous, ranging from benign lesions and indolent renal cell carcinoma (RCC) to highly aggressive phenotypes. Accurately determining tumor aggressiveness before treatment remains a major clinical challenge. Conventional MR imaging lacks sensitivity to microarchitectural differences underlying tumor biology, and biopsy is limited by sampling error and tumor heterogeneity. Radiomics offers a noninvasive approach to extract high dimensional features that may capture biologically relevant patterns. In this IRB-approved retrospective study, we evaluated MR-based radiomics combined with machine learning (ML) to differentiate indolent from aggressive tumors. Pre-operative clinically acquired MR images obtained from eighty-two RCC patients (aggressive, n = 56; indolent, n = 26) who underwent nephrectomy were included for the analysis. Regions of interest (ROIs) of the entire tumor were manually drawn using 3D Slicer on the morphological T2-weighted images and four phases (pre-contrast [PRE], corticomedullary [CM], nephrographic [NG], and delayed [DEL]) from the contrast-enhanced MR images, normalized, and used for radiomic feature extraction (PyRadiomics, v3.1.0). Independently, tumors were grouped into aggressive and indolent phenotypes based on the established histological classification of Mayo Clinic. Statistical analysis used two-sided t-test. Features selected based on mutual information were used for ML classification (Random Under-Sampling Boosting (RUSBoost)) and performance assessed using accuracy, sensitivity, precision, specificity, F1-score, and area under the curve (AUC). Of 107 radiomics features extracted, the following numbers showed statistically significant differences between aggressive and indolent tumors (p < 0.05): CM phase (n = 19), NG phase (n = 24), DEL phase (n = 22), and T2-weighted images (n = 29). Several features consistently differed among contrast-enhanced images and included first-order (length, surface area, volume), GLCM (correlation, IDM, IMC1, IMC2), GLDM (dependence entropy, dependence non-uniformity, dependence variance), GLRLM (gray-level non-uniformity, run entropy, run-length non-uniformity), GLSZM (high gray-level emphasis, zone entropy), and NGTDM (coarseness). Using the 15 features with highest mutual information across 50 patients (aggressive, n = 31; indolent, n = 19), a five-fold cross-validation RUSBoost classifier achieved accuracy 0.81, sensitivity 0.84, precision 0.87, specificity 0.77, F1-score 0.85, and AUC 0.80 for identifying the aggressive phenotype. MRI radiomics features show potential to differentiate aggressive from indolent renal tumors. ML-based classification using these features demonstrates promising accuracy. Further optimization and validation in larger cohorts, may help guide individualized care, including consideration of active surveillance for indolent renal tumors.