Estimating Metastatic Disease Heterogeneity Using Radiomics-based Lesion Clustering and Heterogeneity Index - Application in Lung Adenocarcinoma Patients and Correlations With Patient Outcomes.

Intra-patient inter-lesion heterogeneity contributes to therapeutic resistance and poor prognosis in metastatic lung adenocarcinoma. However, non-invasive approaches to quantify this heterogeneity across multiple lesions remain limited. In this retrospective single-center study, we analyzed 335 patients with newly-diagnosed metastatic lung adenocarcinoma and 1,583 CT-segmentable lesions across two temporal cohorts. Cohort 1 (n = 167; late 2016 to late 2019) served as the training set, and cohort 2 (n = 168; late 2019 to early 2023) as a temporally independent validation set. Radiomic features (n = 68 robust features selected for inter-segmentation reproducibility) were extracted from all lesions and preprocessed using Yeo-Johnson transformation, robust scaling, and rank transformation. Lesions were grouped by unsupervised consensus clustering (partitioning around medoids, Spearman distance). Cluster labels were assigned in validation set using a nearest-centroid approach. Intra-patient heterogeneity was quantified using the Hill number, defined as the effective number of equally abundant lesion clusters per patient. Associations between the Hill number and objective response rate and overall survival were assessed using logistic and Cox regression models. Twelve stable lesion clusters were identified. Higher Hill number was associated with lower objective response rate (adjusted odds ratio = 0.68 [0.47-0.93], p = 0.0237) and shorter overall survival (adjusted hazard ratio = 1.23 [1.03-1.47], p = 0.0243). In the temporally independent validation cohort, adding Hill number improved objective response rate prediction (area under the receiver operating characteristic curve: 0.72 to 0.74). This study proposes an innovative non-invasive radiomics-based framework to quantify intra-patient heterogeneity, with potential value for response prediction in metastatic lung adenocarcinoma.