2585 - Tumour Texture Analysis Using Radiomics for Predicting Treatment Outcomes in Oropharyngeal Cancer

Purpose/Objective(s): Tumours appearing visually similar on contrast-enhanced computed tomography (CECT) may harbor subtle differences imperceptible to the human eye. Radiomics converts medical images into quantitative, mineable data by analyzing voxel-level intensity and texture, capturing tumor heterogeneity, morphology, and spatial complexity. This study analyzes pre- and post-treatment CECT images to correlate radiomic features with tumor behavior and treatment response in oropharyngeal cancer.

Materials/Methods: Patients with biopsy-proven carcinoma oropharynx treated with definitive radiotherapy, with or without chemotherapy, were included. CECT scans were acquired at baseline (radiotherapy planning) and at 3 months post-treatment using a Siemens Biograph mCT S(64)-3R scanner with 3-mm slice thickness. Gross tumor volume was manually delineated to define the Region of Interest. Treatment response was assessed using RECIST 1.1 criteria. Pre- and post-treatment DICOM images underwent radiomic feature extraction using the PyRadiomics algorithm in 3D Slicer (version 5.6.2). 851 features were extracted including 107 original features (shape, first-order, and texture based) and 744 wavelet-transformed features generated using default high- and low-pass filters across three dimensions. Feature extraction and definitions were compliant with PyRadiomics documentation and the Imaging Biomarker Standardization Initiative guidelines.

Results: 90 patients were analyzed with median age of 60 years (range 31–85), male predominance (93%) and mostly locoregionally advanced disease, predominantly Stage IVa (54%) followed by Stage III (30%). Base of tongue was the commonest subsite (63.3%), followed by tonsil (23.2%). 76 patients achieved complete response (CR) at 3 months, and showed 50 common radiomic features characterized by homogeneous intensity distribution, high gray-level emphasis, and organized texture patterns, suggesting viable, well-vascularized tumors with uniform cellularity. 9 partial responders shared 8 variance-based and cluster-related features reflecting increased textural variability, consistent with mixed regions of viable tumor, necrosis, and fibrosis showing differential treatment response. 5 patients with progressive disease exhibited 17 common features marked by large low-intensity regions, marked heterogeneity, and fluctuating signal intensities, suggestive of hypoxic, necrotic, and chaotic tumor architecture. After a median follow-up of 21.6 months, among patients achieving CR, 56 remained disease-free, 10 developed distant metastases, and regional and local recurrences occurred in 6 and 4 patients, respectively.

Conclusion: Radiomic features demonstrate potential to non-invasively characterize tumor microenvironment and predict treatment response. These findings support the integration of radiomics into artificial intelligence–based predictive models to enable personalized treatment strategies in oncology.