ESTRO 2024 - Abstract Book

S4429

Physics - Machine learning models and clinical applications

ESTRO 2024

Keywords: DL auto-planning, Model adaptation, Breast cancer

686

Digital Poster

Predicting Early Toxicities in HN Cancer Patients through CBCT for Personalized Radiation Therapy

Manju Sharma 1 , Jason W Chan 1 , Robert Jeraj 2 , Sue S Yom 1

1 University of California, San Francisco, Radiation Oncology, San Francisco, USA. 2 University of Wisconsin, Medical Physics, Madison, USA

Purpose/Objective:

Early and late toxicities are well-recognized as treatment-limiting factors in head and neck radiotherapy (HNRT) 1 . While considerable attention has been directed to reducing late toxicities to improve patients’ quality of life, early toxicities remain a controversial domain without reproducible and generalizable preventive measures, despite their repercussions on the precision of radiation delivery. This is particularly critical in HNRT, where radiation therapy is typically administered in daily fractions over the course of 6 to 7 weeks. This study aimed to develop a cost and time effective daily CBCT-based prediction model for early toxicities with the objective of optimizing resource allocation by minimizing needless plan adaptations and introducing novel methods to flag when real-time adjustments are needed.

Material/Methods:

In this retrospective study, we analyzed 32 head and neck (HN) cancer patients with ~1000 daily cone-beam computed tomography (CBCT) images. Physicians had graded 25 different early toxicities weekly, according to CTCAE v5. The prescribed radiation doses to the patients varied based on the risk group. The high-risk PTV received 69.96 Gy for definitive treatment without surgery (HR1) and 66.0 Gy for the post-operative approach (HR2). Meanwhile, the intermediate-risk (IR) PTV was prescribed 59.4 Gy, and the low-risk (LR) PTV received 54.12 Gy. These doses were administered in 33 fractions across PTV HR1 , PTV HR2 , PTV IR and PTV LR , respectively. CBCT-based organ-specific daily dose and volume calculations were performed using the ANatomically CONstrained Deformation Algorithm (ANACONDA) in the RayStation Treatment Planning System (TPS) 2 . Gamma evaluations were performed in the Varian MobiusCB module using a (0.2 g/cc/1 mm) criteria. For each patient, we analyzed daily changes in volume and dose metrics for target volumes (specifically, D 95 and D 98 for GTV p , GTV n , GTV T , PTV/CTV HR1 , PTV/CTV HR2 , and PTV/CTV IR ) and organs at risk (OARs) i.e., mean doses for the parotid glands, oral cavity, mandible, and esophagus, along with 0.03 cc doses for the skin, constrictor muscle, spinal cord, and mandible. An Augmented Dickey-Fuller (ADF) test was performed to screen for toxicities associated with specific events, setting a cut-off of at least ten patients with non-stationary changes. Using an Elastic Net model, an in-house Python code was developed to correlate the daily CBCT-based changes with the weekly toxicity changes.

Results: