Abstracts

RETURN TO ABSTRACT LISTING


1895. CT-Based Radiomics Analysis for the Prediction of Tumor Grade in Clear Cell Renal Cell Carcinoma
Authors * Denotes Presenting Author
  1. Natalie Demirjian *; Keck School of Medicine of the University of Southern California
  2. Brandon Fields; Keck School of Medicine of the University of Southern California
  3. Steven Cen; Keck School of Medicine of the University of Southern California
  4. Xiaomeng Lei; Keck School of Medicine of the University of Southern California
  5. Darryl Hwang; Keck School of Medicine of the University of Southern California
  6. Marielena Rivas; Keck School of Medicine of the University of Southern California
  7. Bino Varghese; Keck School of Medicine of the University of Southern California
Objective:
Tumor grade is an independent prognostic indicator in clear cell renal cell carcinoma (ccRCC) and principally guides management. We evaluated the utility of CT-based radiomics data in preoperatively classifying low- (ISUP I-II) and high-grade (ISUP III-IV) ccRCC using a machine learning approach.

Materials and Methods:
203 subjects (mean age 61 years; range 27–85 years) with histologically proven ccRCC were identified as having obtained a preoperative multi-phase CECT of the abdomen and pelvis. 78 scans were acquired from our institution by mining a prospectively maintained server between June 2009 to April 2016, and 125 scans were acquired from the TCGA-KIRC database. The image datasets were loaded into Synapse 3D and manually segmented with oversight from an experienced abdominal radiologist. Whole kidneys were used for image registration across different CECT phases. Post-registration, tumor volume in the nephrographic phase was used as the reference template for the remaining CECT phases. 1708 radiomics features across 9 texture families were extracted using custom data processing algorithms. Independent t-test or Wilcoxon sum rank test with Benjamini and Hochberg correction for multiple testing were used for the descriptive analyses. Random Forest (RF) and Real Adaboost were used with a 10-fold cross-validation to construct the decision classifier. Accuracies of the models were quantified by AUC. Out-of-bag Gini index was used to rank variables of importance.

Results:
Univariate analysis showed strong differences in the values of the radiomic metrics between low- and high-grade groups. Notably, 62.42% were significant at p=0.05. After Benjamini and Hochberg correction, 12.4% of the metrics remained significant. AUCs for RF and Adaboost to correctly classify low- and high-grade ccRCC were 0.66 and 0.60, respectively, indicative of moderate discriminative power. However, 38.62% of the univariate results matched criteria for variable of importance derived from the machine-learning models, particularly GLDM3D.

Conclusion:
Clinical decision making in ccRCC is predicated on an accurate assessment of tumor grade. While low-grade tumors can frequently be managed with close-observation and nephron-sparing surgery, high-grade tumors often necessitate surgical excision. The Fuhrman grading system, though highly predictive of biological aggressiveness and metastatic potential, requires tissue for histopathological analysis. Machine-learning augmented CT-based radiomics analysis has the potential to differentiate between low- and high-grade ccRCC, though validation studies using larger sample sizes are warranted. Evaluating tumor grade pre-therapy is useful both for risk stratification and for treatment planning; machine-learning-based radiomics may provide a feasible evaluation tool once validated.