Purpose: Radiomics machine learning models are a way to analyze existing imaging data to derive clinically relevant information that can be applied to a customized treatment paradigm. This study investigates the use of multi-modality (PET/CT) radiomics based machine learning to predict recurrence in head and neck cancer patients.
Methods: Head and neck contrast enhanced radiation CT simulation scans, PET scans, radiation clinical target volumes (CTV), and clinical outcomes of 201 patients were obtained from The Cancer Imaging Archive (TCIA). Pyradiomics was used to extract intensity, shape, and texture based radiomics features (110) from CT and PET scans separately for all patients. Of the 220 (PET+CT) features extracted, the top 7 performing features were obtained by sequential feature selection in order to form a model for recurrence prediction. Multi-modality and single modality manual feature selection models, based on individual feature AUROC values, were run on the dataset for comparison.
Results: The most accurate model was a multi-modality (PET/CT) model resulting from a combination of 5 PET and 2 CT features selected by backward sequential feature selection with a logistic regression estimator and five-fold cross validation. This model achieved the largest AUROC value (0.7895) compared to multi-modality (0.7701) and single modality (0.7364-CT, 0.7355-PET) manual feature selection models.
Conclusion: Multi-modality radiomics results in more accurate machine learning models for recurrence prediction in head and neck cancer than single modality models. Multi-modality radiomics based recurrence prediction has the potential to improve the clinical outcome of head and neck cancer patients by applying the recurrence risk to pre-treatment management and surveillance patterns.
IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)