Integrating deep learning network and liquid biopsy for the evaluation and prediction of response in patients with advanced non-small cell lung cancer: Initial results from the RATIONALE prospective study.

e20036 Background: Lung cancerstill represents one of the most common and lethal diseases worldwide. There remains an unmet clinical need for the discovery of biomarkers to refine the management of patients with advanced non-small cell lung cancer (NSCLC). The combined evaluation of circulating cell-free DNA (cfDNA) and deep learning-based radiomic methods remains to be validated in the clinical setting. Methods: The RATIONALE study is a retro-prospective observational clinical trial consecutively enrolling treatment-naïve patients with advanced NSCLC who received standard first-line systemic treatments at Paolo Giaccone University Hospital, Palermo (Italy). Baseline and follow-up blood samples were collected according to computed tomography (CT) examinations performed using the same equipment (Siemens SOMATOM 128 slices), the same acquisition protocol, and the same contrast medium (Iomeron 400) every twelve weeks. All primary lung tumor internal regions of interest (ROIs) were segmented. Radiomic feature extraction was performed on the ROIs using Pyradiomics, an extension of the 3DSlicer. Circulating cfDNA was extracted from 1 ml of plasma using a QIAamp Circulating Nucleic Acid Kit (Qiagen) and quantified in terms of ng/μL using a QubitTM dsDNA HS Assay Kit (ThermoFisher Scientific). The diagnostic performance was evaluated using receiver operating characteristic (ROC) curves. Results: Between June 2021 and November 2023, fifty-five patients were included. Here we present the first results from the cohort of twenty patients undergoing first-line pembrolizumab-based treatments. Among the thirteen features extracted, the feature 'original_glcm_Id' evaluating the local image homogeneity emerged as statistically significant for differentiating adenocarcinoma and squamous cell carcinoma (sensitivity of 83.7%, specificity of 81.0%, accuracy of 82.8%; p < 0.001). When correlating cfDNA dynamics and radiologic response, out of the eight features examined, the feature 'first order_mean' measuring the average intensity of gray levels proved to be statistically significant (sensitivity of 86.0%, specificity of 87.7%, accuracy of 86.9%; p < 0.006). Finally, the feature 'size zone non-uniformity' measuring the variability of volumes in different gray level zones within the image showed a good performance in predicting overall survival (sensitivity of 65.8%, specificity of 100.0%, accuracy of 77.8%; p < 0.001). Conclusions: The preliminary results of this ongoing study suggest the combined evaluation of liquid biopsy and radiomic features for obtaining minimally invasive and cost-effective biomarkers in the first-line clinical setting of NSCLC undergoing first-line pembrolizumab.