Preoperative prediction of high‐risk lymph nodes in penile cancer: Graafland criteria vs radiomics

Penile squamous cell carcinoma is a rare cancer although the incidence varies across the globe according to geographic differences in risk factors [1]. The primary site of lymphatic metastatic disease is the groin and early surgical staging of the groins offers a survival advantage [1, 2]. There are certain histopathological features of nodal metastatic disease that indicate high-risk disease with worse prognosis: extra-nodal tumour extension, three or more unilateral metastatic nodes, and pelvic lymph node involvement [3, 4]. These patients could benefit from pre-surgical neoadjuvant therapy, but the features are not detected until histopathological analysis after surgery. In 2011, Graafland et al. [3] showed that if inguinal lymph nodes showed necrosis or irregular margins on preoperative CT scans this could predict high-risk penile cancer with sensitivity of 95% and specificity of 82%. This creates the opportunity to offer patients with high-risk disease neoadjuvant therapy prior to surgery. The Graafland criteria are currently used in clinical practice but have never been validated beyond the initial Graafland et al. [3] small series. Radiomics is the quantitative evaluation of regions of interest on radiological imaging studies that reflect size, shape, density, and texture [5]. The goal of radiomics is to extract clinically meaningful information about a segmented tumour or region of interest that is not readily detectable by the human eye [5, 6]. Radiomic analysis of medical imaging can be used to predict or correlate with different clinical outcomes. For example, radiomic analysis of MRI and CT scans has been used to grade bladder tumours and to predict metastatic nodes in bladder cancer [7]. Our group undertook a project to reproduce the Graafland criteria to predict high-risk nodal disease in penile cancer and evaluate the potential of radiomics to risk stratify these cases in a more objective manner. We performed a retrospective chart review to identify 38 patients who were treated between 2016 and 2020 with histologically confirmed penile squamous cell carcinoma and CT imaging completed prior to lymph node dissection. Patients were excluded if they had previously undergone any systemic chemotherapy for penile carcinoma or any previous radiotherapy to the inguinal or pelvic lymph nodes. No other restrictive inclusion criteria were applied. Patients were staged according to the American Joint Committee on Cancer (AJCC) TNM staging system, eighth edition. All CT scans were performed using intravenous contrast with slice thickness of 3 mm. Two experienced oncological radiologists (B.M. and A.A.) applied the Graafland criteria to the inguinal lymph nodes to record a score of low risk or high risk for every CT examination and this was compared to the pathological presence of high-risk features at lymphadenectomy. For the radiomics analysis, the largest inguinal lymph node was segmented on each patient's CT scan to generate a volume of interest (VOI) using a semi-automated approach with Healthmyne Software (Genentech Roche, San Francisco, CA, USA). One patient was excluded from the analysis secondary to incompatibility of the CT scan with the radiomics software so the radiomics sample size was 37. A total of 308 radiomics features were extracted using the VOIs and tested for predicting high-risk disease at subsequent lymphadenectomy surgery. Among 308 radiomics features, two features were non-informative, and thus they were excluded from the analysis. The features with poor predictive value (i.e., area under the curve [AUC] <0.5) were also 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. Thus, a total of 109 features were screened out to address the collinearity issue, and predictive models were developed based on the remaining 199 features. The principal component analysis (PCA) method and t-distributed stochastic neighbour embedding (t-SNE) were used for dimension reduction. The support vector machines (SVMs) with Gaussian and linear kernels were considered after the dimension reduction. Two logistic regression models were developed using stepwise selection and least absolute shrinkage and selection operator (LASSO) approach. The single decision tree (DT) was considered. As the SVM with Gaussian kernel was deemed to be overfitted, it was excluded. The performance of these five predictive models was compared by using the five-fold cross-validation (CV) method with 100 replicates. A total of 38 patients were included in the study. In all, 16 patients (42%) had postoperative pathology with high-risk lymph node disease. Out of the 38 scans reviewed by the radiologists, the concordance rate of assignment of high-risk disease was 89.5% (34/38). Results for the radiologist's assessment using the Graafland criteria yielded results of: accuracy: 33/38 = 86.8% (95% CI 71.9–95.6%), AUC: 0.869 (95% CI 0.758–0.981), sensitivity = 14/16 = 87.5%, and specificity = 19/22 = 86.4%. The performance of the five predictive radiomic models were compared by using the five-fold CV method: logistic with stepwise selection (P-enter = 0.05 and P-stay = 0.05; F45 + F279), logistic with LASSO (F53 + F150 + F45 + F279), PCA + SVM with linear kernel (44.8% explained by the first two PCs), t-SNE + SVM with linear kernel, and the single DT (Table 1). The single DT looked best in terms of AUC values in the training procedure. However, based on the five-fold CV with 100 replicates results, the stepwise selection model seems best among them, and it is transparent and parsimonious. The single DT seems to be relatively poor as it may be overfitted. The stepwise model yielded a diagnostic accuracy/AUC of 0.929 with a five-fold CV value of 0.921. The results of this study (sensitivity = 87.5% and specificity = 86.4%) were able to replicate the good results of Graafland's original work (sensitivity = 95% and specificity = 82%). This is a significant and reassuring finding since these criteria are currently used in practice. Although the Graafland criteria were validated in this retrospective setting, there are challenges in ‘real-word’ application. It is questionable if these results could be reproduced outside of academic centres with subspecialised radiologists who deal frequently with penile cancer. There is also the potential of inter-reader variability. It is difficult to distinguish necrosis from intra-nodal fat and extra-nodal extension/irregular borders can be highly subjective. The radiologists in this study disagreed about the presence of high-risk features in four out of the 38 cases (10.5%), which highlights the subjective nature of the criteria. The radiomics results were also promising with an AUC of our potential model of 0.921 ± 0.096 (±SD), slightly better than expert radiologist analysis (AUC: 0.869, 95% CI 0.758–0.981). Our study shows that radiomics has the potential to preoperatively detect high-risk nodal disease objectively and in a standardised manner that does not depend on radiologist expertise. Although this work shows the promise of radiomics, there are several limitations. This study was retrospective in design and direct correlation between pathology and radiomic analysis of specific lymph nodes was not possible. To overcome this limitation, the largest lymph node on each CT scan was selected for radiomic analysis. It is highly likely that when high-risk nodal disease was present, the largest lymph node would possess these features, so the assumption is reasonable. Additionally, a larger sample size would be useful to both train and further optimise our model, ideally utilising an external validation data set. Despite these limitations, this does show the potential of radiomics to risk stratify nodal disease in penile cancer without the subjectivity of human analysis or requiring extensive expertise. The authors declare that they have no conflicts of interest.