ESTRO 2024 - Abstract Book

S1662

Clinical - Lung

ESTRO 2024

Milan, Italy. 4 University of Piemonte Orientale (UPO), Department of Translational Medicine, Novara, Italy. 5 European Institute of Oncology IRCCS, Radiation Research Unit Oncology, Milan, Italy. 6 European Institute of Oncology IRCCS, Scientific Directorate, Milan, Italy

Purpose/Objective:

Patients diagnosed with early-stage non-small cell lung cancer (ES-NSCLC) are often candidate to curative-intent SBRT when unfit to surgery. It is not rare that their comorbidities contraindicate a histopathological assessment, as well: this is a well-known issue in thoracic Radiation Oncology. Such clinical challenge results into the still unmet need of identifying non-invasive biomarkers of disease, to better discriminate benign vs non-malignant lesions, and to attempt a characterization of the most suspected lesions. Albeit promising, radiomics has not led to the implementation of any signature in this clinical setting. This study aims to assess the role of image filtering in the construction of (clinico)-radiomic models to predict oncological outcomes on a homogenous cohort from a tertiary care cancer center.

Material/Methods:

The institutional dataset was screened per the following inclusion criteria: 1) adult patients with radiological and/or histopathological diagnosis of ES-NSCLC according to the AJCC TNM 8th edition, 2) curative-intent SBRT to a minimum biological effective dose- BED- of 100 Gy (assuming an α/β ratio of 10 Gy), 3) availability CT simulation scans, 4) minimum follow-up of at least 12 months since diagnosis and 5) availability of a written informed consent for use of data for clinical research and educational purposes. Segmentations were manually performed by a single Radiation Oncologist to overcome inter-observer variability. The outcomes of interest were: overall survival (OS), progression-free survival (PFS), and local progression-free survival (LPFS). The open-source IBSI (Imaging Biomarker Standardization Initiative)-compliant Pyradiomics software (v3.0.1) was used for both image pre-processing and radiomic features (RFs) extraction. All in-built filters were enabled, including all nine permutations of the wavelet filter and subcategories of the lbp-3D image type. Following the exclusion of RFs with zero variance and high correlation (Spearman ρ > 0.95), an iterative clustering algorithm was used to group RFs with Spearman ρ > 0.75; subsequently, a multivariable Cox — LASSO Regression Model was applied on each RF selected in the previous step. Univariate Cox PH regression models were used to test associations between clinical variables and the radiomic score; variables with p ≤ 0.10 in univariate analysis were included in multivariate analysis and retained if the p -value was confirmed as ≤0.10. All model performances are expressed in terms of AUC (Area Under the Curve), for the test set.

Results:

A total of one hundred patients treated between September 2013 and July 2022 were included in the analysis. Median age at diagnosis was 76 (IQR: 70-82) years, with nearly all patients (n=92) presenting with at least one comorbidity. Median Charlson Comorbidity Index was 7 (IQR: 6-8); and 78 patients were identified as being either active or former smokers. The most frequently represented stages were IA2 (n=46) and IB (n= 21). Overall, 54 patients underwent histopathological assessment, which resulted adenocarcinoma in 41 cases, and of squamous cell carcinoma in the remaining 13 cases. Considering clinical variables, the BED was significantly associated to OS (p=0.02) at multivariable analysis, while FEV1% was selected as a predictor for PFS and LPFS (p=0.02 and p=0.019, respectively). For the radiomic model, the LASSO mostly selected features of the higher-order statistics category