<sup>18</sup>F-FDG PET/CT Habitat Radiomics for Predicting Epidermal Growth Factor Receptor Mutation Status in Locally Advanced Non-small Cell Lung Cancer: A Multicenter Study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Nuclear Medicine, Siping Central People's Hospital, Siping, Jilin, China.
- Department of Radiology, Siping Central People's Hospital, Siping, Jilin, China.
- Center for Nuclear Medicine and Molecular Imaging, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China.
- Department of Radiology, The Second Qilu Hospital of Shandong University, Jinan, Shandong, China.
Abstract
IntroductionAccurate identification of epidermal growth factor receptor mutation status is essential for treatment planning in locally advanced non-small cell lung cancer, but tissue testing may be limited by specimen availability and sampling. We developed and externally validated a fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography habitat radiomics model to estimate mutation probability as an adjunct to molecular testing.MethodsThis retrospective multicenter study included 313 patients with locally advanced non-small cell lung cancer who underwent imaging and tissue-based mutation testing. Center A provided the training cohort (n = 142) and internal validation cohort (n = 61), and Center B provided the external validation cohort (n = 110). Separate Otsu thresholds were applied to positron emission tomography and computed tomography intensities within each segmented tumor. Voxel-wise intersection generated four mutually exclusive habitats. Habitat features were selected in the training cohort and evaluated with six classifiers. Additional analyses examined adenocarcinoma-restricted performance, alternative partitioning, segmentation reproducibility, calibration, decision-curve performance, and selection bias.ResultsThe training, internal validation, and external validation cohorts included 69/73, 31/30, and 64/46 mutation-positive/mutation-negative tumors, respectively. The support vector machine habitat model achieved areas under the receiver operating characteristic curve of 0.911 (95% confidence interval, 0.869-0.948), 0.898 (0.819-0.967), and 0.884 (0.829-0.933), respectively. With the threshold selected only in the training cohort, external validation sensitivity, specificity, accuracy, and negative predictive value were 0.750, 0.739, 0.745, and 0.680. The combined model achieved an external validation area under the curve of 0.904 (0.855-0.949).ConclusionHabitat radiomics showed promising discrimination across two centers, including among patients with adenocarcinoma. The model may help prioritize confirmatory testing but should not replace biopsy, plasma testing, or molecular analysis. Prospective validation and direct biological correlation are required before clinical implementation.