Feasibility of radiomics and deep learning methods for prognosis of non–small cell lung cancer using 18F-FDG PET/CT

Résumé fourni par la source

Accurate prognostication for non-small cell lung cancer (NSCLC) is challenging due to tumor heterogeneity and the limitations of conventional PET biomarkers. This technical feasibility study compared the difference in prognostic performance of radiomics and deep learning (CNN) methods against conventional biomarkers such as SUVmax and TLG to predict 1-year overall survival (OS 1 ) in 150 treatment-naïve patients. Using automated lesion segmentation, internal results showed AUCs between 0.75 and 0.80, while external validation initially yielded AUCs from 0.58 to 0.68. Following bootstrapped subsampling, all models showed improved performance, led by the TLG-only model with an AUC of 0.83. Study scope was limited by the lack of feature-level harmonization and variations in image quality and segmentation. Although simple PET biomarkers performed comparably to more complex architectures, they offer the advantage of being robust and clinically accessible. In summary, while radiomics and deep learning models each show feasibility for PET-based prognostication in NSCLC, the radiomics model currently provides greater stability and interpretability, and TLG continues to represent an effective, clinically practical biomarker.