Abstract PO-043: Development of an automated multi-objective model utilizing delta radiomics to predict locoregional recurrence in head and neck cancer patients treated with primary radiation

Abstract Head and neck squamous cell carcinoma (HNSCC) is the sixth most common cancer globally and is frequently treated with definitive radiotherapy. However, 15-50% of patients experience locoregional recurrence (LRR) within 3 years of treatment. The majority of recurrences develop within the first 2 years. Machine learning and deep learning have been utilized to predict oncologic treatment outcomes with high accuracy. Accurate prediction allows for the identification of high-risk patients prior to recurrence, allowing for more intensified monitoring and increased likelihood of long-term disease control. Radiomics is the extraction of quantitative data from radiologic images and provides information that is imperceptible to the human eye. Prior studies have successfully utilized pre- or post-treatment contrast enhanced CT (CECT) images to predict treatment outcomes in HNSCC in machine learning models. In this study, we utilize pre-treatment CECT radiomic features to build an automated multi-objective (AutoMO) model to predict LRR for patients with HNSCC treated with definitive radiotherapy. Patients with primary laryngeal, oropharyngeal, or hypopharyngeal cancer treated with primary radiotherapy from January 2018 through November 2020 were identified. Patients were excluded if they did not have a pre-treatment CECT within 3 months of initiation or a post-treatment CECT within 1-6 months of completion of radiotherapy. Pre- and post-treatment CECT scans as well as clinical information, including locoregional recurrence or progression within two years, were collected retrospectively. Pre-treatment CECT scans contoured for gross tumor volume (GTV) were also collected, which were utilized as the ground-truth reference for auto-contouring of pre-treatment scans. An AutoMO model was developed that maximizes both sensitivity and specificity simultaneously, offering more balanced results as compared to a model that reports the traditional area under the curve (AUC) metric. Following CECT image pre-processing, 275 radiomic features including intensity, texture and geometry were extracted in 3-dimensional volume from pre-treatment CECT scans. Image features were fed into the AutoMO model. 3-fold cross validation was performed, and hyperparameters were tuned to optimize model results. A total of 70 patients met inclusion criteria. The population was 81.4% male (n=57) with an average age of 64.4. Of all patients 22 (31.4%) were positive for recurrence or progression. The model achieved a sensitivity of 0.55, specificity of 0.83, accuracy of 0.79, and AUC of 0.65. The AutoMO model was moderately accurate for predicting recurrence or progression for patients with HNSCC utilizing pre-treatment CECT radiomics. Future iterations integrating clinical data, post-treatment images, and an increased sample size should improve model accuracy. Citation Format: Bryan Renslo, Ethan Kallenberger, Patrick Ioerger, Kenny Guida, Oluwatobiloba Ige, Omar Karadaghy, Gregory Gan, Zhiguo Zhou, Andres Bur. Development of an automated multi-objective model utilizing delta radiomics to predict locoregional recurrence in head and neck cancer patients treated with primary radiation [abstract]. In: Proceedings of the AACR-AHNS Head and Neck Cancer Conference: Innovating through Basic, Clinical, and Translational Research; 2023 Jul 7-8; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2023;29(18_Suppl):Abstract nr PO-043.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.