ESTRO 2024 - Abstract Book

S5135

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

1 Trinity College Dublin, Applied Radiation Therapy Trinity (ARTT), Discipline of Radiation Therapy, School of Medicine, Trinity St. James's Cancer Institute, Dublin, Ireland. 2 Institut Gustave Roussy, Université Paris-Saclay, Institut Gustave Roussy, Inserm, Radiothérapie Moléculaire et Innovation Thérapeutique, Villejuif, France

Purpose/Objective:

The prediction of xerostomia is limited by the inability of traditional NTCP models to account for the complexity of the parotid gland’s (PG) architecture and it’s response to radiation. Quantitative imaging analysis, through radiomic feature (RF) extraction, can document the complexities of PG anatomy and may provide previously “unseen” imaging biomarkers to improve the accuracy of predictive models. This study aimed to improve the prediction of acute xerostomia and sticky saliva (SS) through the inclusion of radiomic analysis of the PG for patients with head and neck cancer (HNC).

Material/Methods:

Anonymous data was collected from 98 patients previously treated for HNC with IMRT to a total dose of 69.96Gy in 33 fractions. Data collected included patient age, sex, diagnosis, disease stage, history of smoking and alcohol consumption, chemotherapy agents, mean dose to the left and right parotid glands, and planning CT images. Baseline xerostomia and SS scores using EORTC QLQ-HN35 questionnaire were collected. One hundred RF were extracted from the patients’ left and right PGs using LifeX (v7.3) with a bin size of 5HU and resampled voxel size of 1x1x1mm. Following confirmation that the ipsilateral and contralateral PG RF values were similar per patient based on Wilcoxon Signed Rank test (p>0.05), the pairs of extracted features per patient were averaged, resulting in a single value per feature for each patient. The clinical outcomes were acute xerostomia and SS measured within 5 months of baseline; dichotomised as no/mild and moderate/severe. Descriptive and univariate analysis provided an overview of patient data and their association with xerostomia and SS. Dimension reduction of the RFs was performed based on principal component analysis (PCA) and three logistic regression models were built based on clinical variables (age, sex, smoking history, alcohol consumption, and mean parotid dose); nine RFs derived from the PCA; and the combination of the clinical and radiomic data. Predictive power was quantified using balanced accuracy and AUC-ROC.

Results:

At baseline, 74 (77%) out of the 96 patients with data available presented with no/mild xerostomia and SS respectively. Based on available data from 65 patients at follow up, 71% experienced moderate/severe acute xerostomia and 69% experienced moderate/severe amounts of SS. Patient age, sex, smoking or alcohol status, or mean dose to the parotid gland was not associated with acute moderate/severe xerostomia or SS. Following listwise deletion due to missing data, the regression analyses were based on 54, 59, and 54 patients for the clinical, radiomics, and combined models respectively. Relying on clinical data and mean dose to the parotids alone to predict acute xerostomia was almost the equivalent of random guessing with an AUC-ROC of 0.64, and a balanced accuracy of 53%. Analysing RF exclusively achieved an AUC-ROC of 0.77 and a balanced accuracy of 61%; however,