Introduction:

The ability to reliably predict benign renal neoplasms from true kidney cancer without adjunctive procedures is currently lacking. Approximately 20% of patients will have benign pathology after surgery. Current guidelines recommend all patients with a suspected renal mass obtain multiphase cross-sectional imaging, and computed tomography (CT) is the most commonly utilized imaging modality. Radiomics refers to the high-throughput extraction of disease-specific data from routine medical images that are not perceivable by the human eye. While the evaluation of renal masses using machine learning-based radiomic protocols has been validated, studies to date have been limited in scope and by the absence of non-radiographic clinical variables in their analysis. Herein, we use an artificial intelligence platform, combining patient clinical factors with radiomics to create a “virtual biopsy” that may serve as a surrogate for tumor histology.

Methods:

Patients were identified through a prospectively maintained database of patients who underwent partial or radical nephrectomy from May 2007-September 2018. An IRB waiver of consent was obtained prior to data collection. All patients included in the study had clinically localized renal masses identified on multiphase CT imaging followed by open or laparoscopic partial or radical nephrectomy. Demographic, clinical and tumor-specific variables were evaluated. CT-based texture analysis (CTTA) and shape analysis were performed, as previously described. Univariate analysis was performed using all clinical and radiomic features. Clinically significant variables were then incorporated into a Random Forest model. A training cohort of 50 patients was validated in the remaining sample of patients. Decision tree analysis was then performed, with bootstrapping at each tree. 10-fold cross validation was used to obtain robust classification accuracy. Area under the curve (AUC) was used to assess robust discrimination power.

Results:

Of 684 included patients, 76.2% had true malignancy, while 23.8 had benign pathology. 68.4% of patients were male. 43.9% were symptomatic on presentation. The probability of malignant disease was calculated for renal tumors < 7 cm (n = 562) using an established nomogram, which had a diagnostic accuracy of 66.2%. Radiomic evaluation through CTTA and shape analysis yielded a probability of malignancy that ranged from 47.5%- 95.1%. A univariate analysis identified age, male gender, smoking status, comorbid diabetes or hypertension, morbid obesity, symptomatic presentation, high clinical nomogram score (³ 90%), and radiomic probability of malignancy (³70%) as significant variables. The combination of clinically-significant variables and radiomic analysis yielded the highest sensitivity, specificity, positive and negative predictive values, and accuracy (Figure 1). AUCs for clinical nomogram, clinical factors, radiomics, and combination of clinical factors and radiomics was 0.5 (CI 0.43-0.58), 0.63 (CI 0.55-0.71), 0.83 (CI 0.77-0.89), and 0.86 (CI 0.8-0.91), respectively.

Conclusion:

The competency of the analysis to predict benign from malignant disease was highest when radiomic and clinical features were combined. 15.5% of patients had asymptomatic benign disease <4 cm. A safe assumption can be made that these masses were removed unnecessarily. Had surgical intervention been based on the combination of clinical and radiomic predictors of malignant disease, the number of unnecessary surgeries would have dropped to 3.5%. Clinical factors alone were weakly predictive and best considered when incorporated into a broader framework of radiomic analysis. These findings suggest that clinico-radiomic artificial intelligence platforms may facilitate improved pre-operative risk stratification. Implications are broad and include an expansion of active surveillance protocols for renal masses, personalized imaging regimens in patients on active surveillance based on their clinico-radiomic risk, and corroboration or adjudication of histological findings in patients undergoing biopsy, all of which imply a considerable cost-savings potential.

Funding: N/A