Application of radiomics as a risk stratification tool in node positive penile cancer.

3 Background: The risk level of lymph node metastasis in penile cancer is determined using the Graafland criteria which determines how patients are treated. These imaging-based criteria have critical implications on patient care yet have not been extensively validated. The purpose of this study was to validate the Graafland criteria and to assess the use of radiomics analysis of CT scans to standardize risk stratification. Methods: Thirty-eight patients with squamous cell penile cancer with CT scans prior to regional lymphadenectomy were included in this retrospective cohort. Patients with prior chemotherapy for PSCC or any inguinal and pelvic radiotherapy were excluded. All CT scans were performed using IV contrast with axial slice thickness of 3mm. CT scans were analyzed by two experienced oncologic radiologists using Graafland criteria. The largest inguinal lymph node in each patient was segmented and radiomics features were extracted and compared to pathologic evidence of high-risk disease. 308 radiomic features were identified. Features with poor predictive value (i.e. AUC <0.5) were excluded. Pearson’s correlation coefficient was used to find the clusters of features (absolute value of correlation of ≥0.95), and the most predictive feature of each cluster was selected.199 features with best predictive value were selected for analysis. Two logistic regressions were performed using stepwise selection and LASSO approach. Results: In 38 subjects, 16 of 38 patients were identified as harboring lymph node metastasis on surgical pathology. Applying Graafland criteria had an inter-reader concordance of 89.5% (34/38), an accuracy of 86.8% (95%CI: 71.9% - 95.6%); AUC: 0.869 (95% CI: 0.758– 0.981); Sensitivity 87.5%; and Specificity 86.4% in predicting high-risk disease. CT radiomics analysis yielded a model including two features F54 (compactness) and F279 (3D wavelet_P1_L2_C1) which is a multi-order texture-based feature. The logistic regression with stepwise selection included two features (F54 and F279) and had an (AUC = 0.921 ± 0.096) after five-fold cross validation. Conclusions: Graafland criteria accurately predicted the presence of high-risk nodal disease on pre-operative CT scans when assessed by expert radiologists. Radiomic analysis shows promise in further refining and standardizing the detection of high-risk nodal disease with the advantage in ease of training and/or clinical application.