Machine learning and magnetic resonance imaging radiomics for predicting human papilloma virus status and prognostic factors in oropharyngeal squamous cell carcinoma

Abstract

Background: We attempted to predict pathological factors and treatment outcomes using machine learning and radiomic features extracted from preoperative magnetic resonance imaging (MRI) of oropharyngeal squamous cell carcinoma (OPSCC) patients.

Methods: The medical records and imaging data of 155 patients who were diagnosed with OPSCC were analyzed retrospectively.

Results: The logistic regression model showed that the area under the receiver operating characteristic curve (AUC) of the model was 0.792 in predicting human papilloma virus (HPV) status. The LightGBM model showed an AUC of 0.8333 in predicting HPV status. The performance of the logistic model in predicting lymphovascular invasion, extracapsular nodal spread, and metastatic lymph nodes showed AUC values of 0.7871, 0.6713, and 0.6638, respectively. In predicting disease recurrence, the LightGBM model showed an AUC of 0.8571. In predicting patient death, the logistic model showed an AUC of 0.8175.

Conclusions: A machine learning model using MRI radiomics showed satisfactory performance in predicting pathologic factors and treatment outcomes of OPSCC patients.