ESTRO 2024 - Abstract Book

S5067

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Rates of locoregional recurrence in oropharyngeal squamous cell carcinoma (OPSCC) are estimated between 9 and 35% despite treatment with curative intent [1,2]. Risk of recurrence remains disproportionately high for patients with additional risk factors, most notably Human Papillomavirus (HPV) negative disease and/or significant smoking history. MRI imaging provides a non-invasive technique for early treatment response assessment during radical (chemo- )radiotherapy (RT). Radiomics, the process of extraction of quantitative features from medical imaging using pre defined mathematical algorithms, can provide high-dimensionality data for inclusion in predictive models. This study aimed to determine if MRI-derived radiomics features (RFs) could predict long-term primary site control in OPSCC.

Material/Methods:

Clinical and imaging data was derived from the MeRInO cohort [3]. Patients with high-risk stage III/IVa OPSCC (AJCC Staging Manual, 7th Edition) were eligible. High-risk was defined as >10 pack year smoking history and/or HPV negative disease. All patients received first-line radical RT 65 Gy in 30 fractions over 6 weeks delivered using VMAT with 6MV photons +/- Cisplatin chemotherapy 100mg/m 2 in weeks 1 and 4, in accordance with institutional and UK national standards. Two research MRI scans were obtained at baseline and in week 3 of RT and imported into the RT planning system (ARIA®, Varian medical systems, Palo Alto, CA) to define the regions of interest. Primary tumour volumes were jointly outlined by an expert head and neck oncologist (CP) and radiologist (IMcC) on the T1 post Gadolinium contrast fat-saturated sequences. The outcome of interest was primary site relapse, defined as histopathologically confirmed oropharyngeal recurrence within 2 years of RT completion. Follow-up imaging was matched with the RT planning CT using rigid registration to confirm the presence of the recurrent lesion was within the 95% isodose. Image pre-processing and RF extraction was performed on the baseline images using the Pyradiomics package [4]. MRI images and associated structure sets were resampled to 1x1x1 mm pixel size using B-spline interpolation. Analysis was performed in R v4.3.1 [5]. RFs were z-score normalised and highly correlated features removed using Pearson’s correlation test, cutoff value 0.9. Univariate associations were assessed using Welch’s t-test, α = 0.05 with Bonferroni correction. Separate penalised ridge regression models using 100 iterations of bootstrap were fitted for clinical and radiomics features. Clinical characteristics for relapse (n = 10) and control (n = 53) groups are presented in the Table. Participants that relapsed were significantly more likely to have high T-stage and low N-stage disease. One hundred and four RFs were extracted, 7 remained following dimensionality reduction and significance testing. Boxplots of non-normalised RF values are presented in Figure 1. Primary tumours which relapsed were significantly larger at baseline (Least axis length (P = 0.009), Maximum 2D diameter (Row) (P = 0.042) and Maximum 2D diameter (Slice) (P = 0.011)) and associated with increased textural heterogeneity (Joint Energy (P 0.015) and Joint Entropy (P = 0.031)). Although absolute differences for the final two features (IDMN and SDLGLE) were small, they were highly statistically significant (P = 0.003 and 0.008 respectively). Figure 2 presents ROC curves for the logistic regression models. Baseline primary MRI RFs demonstrated better capability for prediction of primary recurrence than a model of T- and N-staging groups, mean AUCs were 0.650 and 0.755 respectively. Results:

Relapse (n = 10)

Control (n = 53)

P

Age (mean (SD))

59.1 (5.2)

56.8 (8.0)

0.380

Male (%)

7 (70)

45 (85)

0.494