Enhancing Radiomics Robustness Using Bayesian Penalized Likelihood Pet Reconstruction: Application to Phantom and Non-Small Cell Lung Cancer Patient Studies Publisher Pubmed



Z Valibeiglou ZAHRA ; J Pirayesh Islamian JALIL ; Y Soleymani YUNUS ; S Farzanehfar SAEED ; F Aghahosseini FARAHNAZ ; N Gilani NEDA ; A Rahmim ARMAN ; P Sheikhzadeh PEYMAN
Source: BMC Medical Imaging Published:2025

Abstract

Purpose: This study aims to enhance the diagnostic and prognostic capabilities of PET imaging through improved robustness of radiomics features, utilizing the Bayesian penalized likelihood (BPL) reconstruction algorithm. Specifically, we focus on 18F-FDG PET imaging of lung cancer, which, with non-small cell lung carcinoma (NSCLC) as its most prevalent form, continues to be a leading cause of cancer-related mortality worldwide. The early detection and precise staging of NSCLC are crucial for effectively managing and treating the disease. Method: We studied a NEMA image quality (IQ) phantom and 15 patient PET lesions (14 NSCLC patients selected from 30 patients originally considered). The study assessed the stability of radiomics features against various imaging parameters, emphasizing the impact of the BPL reconstruction algorithm with varying β-values (50, 100, 150, 200, 250, 300, 350, 400, 450, 500, 600, and 700) and three phantom lesion to background ratios (LBRs) of 2:1, 4:1, and 8:1. Manual segmentation was performed, and subsequently, 130 radiomic features were extracted from the reconstructed images. The stability of radiomics features was assessed by calculating the coefficient of variation (COV) for each feature across variations in reconstruction parameters. A COV of ≤ 5% indicated high stability. Results: Our results indicate that morphological and intensity-based features exhibit excellent stability, with a COV of less than 5%. Texture-based features, despite their complexity, also demonstrated robustness. Specifically, 32.3%, 39.2%, 42.3%, and 37.6% of features exhibited high stability in phantom LBR 2:1, phantom LBR 4:1, phantom LBR 8:1, and patient studies, respectively. Overall, 13 morphological, 8 intensity, 6 intensity-histogram, and 5 texture-based features were found to be highly stable against different LBRs and reconstruction parameters. Conclusions: The BPL reconstruction algorithm may enhance the robustness of PET radiomics features, supporting their use in clinical settings for non-invasive diagnosis and staging. The adoption of BPL towards improved PET radiomics robustness has the potential to transform NSCLC evaluation and management, but still needs standardization. © 2025 Elsevier B.V., All rights reserved.