OBJECTIVE: To compare radiomics models of treatment response in oropharyngeal squamous cell carcinoma (OPSCC) patients using features derived from the primary tumour (PT), the metastatic lymph nodes (LNs), and a combination of both (PT+LN). We hypothesize that inclusion of metastatic lymph node features may yield improved performance.

METHODS: Ethics approval was obtained. Eighty-four patients who underwent curative treatment for OPSCC with chemo/bio/radiotherapy were included retrospectively from 2006–2016. The primary tumour and one target lymph node (deemed the most radiologically pathological) were segmented from pre-treatment CECT scans (512x512 matrix, 1–3 mm slice thickness, 120–140 kVP) in two dimensions. The pyradiomics tool was used to extract 760 first, second and third-order features in the native and wavelet domains. Analysis was performed in Matlab. Feature selection was based on Spearman's rank correlation to response. Five-fold cross validation was used. Ten different machine-learning classifiers were trained with 1-to-10 features, obtaining 100 models each for the PT, LN, and PT+LN groups.

RESULTS: AUC was computed for each model. The Wilcoxon rank sum test, using Bonferroni correction, showed a statistically significant increase in AUC for the PT+LN models (p=4.78x10e-4) overall: specifically, 4-of-10 models showed improved performance when trained on combined primary tumour and lymph node features. There was no significant difference in AUC between the PT- and LN-only models (p=0.49).

CONCLUSION: Our preliminary results suggest that prognostic radiomics models in OPSCC can use metastatic lymph node features to further improve their performance. As well, there may be a role for lymph node-only analysis in cases of occult primaries.

Radiomics Prediction Models of Oropharyngeal SCC Recurrence: Improving Performance with Metastatic Lymph Node Features