Introduction
Three-dimensional tumor spheroids exhibit dynamic collective behaviors, including spontaneous reorganization, migration dynamics, and epithelial-mesenchymal transition, observable through label-free time-lapse microscopy. However, standard statistical methods fail to account for temporal dependencies inherent in longitudinal imaging data, treating autocorrelated timepoints as independent observations. This methodological limitation inflates effective sample sizes approximately 4-fold, yields anticonservative hypothesis tests, and compromises the reliability of spheroid-based drug screening applications.
Methods
We developed a validated computational framework integrating multi-scale texture analysis (37 features: GLCM, wavelet, Gabor descriptors) with autocorrelation-informed statistical testing. The framework incorporates: (1) global standardization addressing 24 orders of magnitude variance across feature types, (2) empirical autocorrelation characterization revealing 4-hour median decorrelation lags, and (3) block bootstrap resampling with 5-hour blocks enabling valid statistical inference. Validation was performed using non-small cell lung cancer spheroids (A549 and H1299 cell lines, 48 hourly observations per condition) treated with capecitabine (50 μM), with external validation using Cancer Dependency Map RNA-sequencing data.
Results
The framework successfully discriminated treatment conditions with texture features, achieving over 100-fold improvement compared to the mean intensity baseline for mesenchymal-phenotype cells. External validation demonstrated quantitative correspondence between texture-derived discrimination ratios and independent molecular markers (VIM/CDH1 expression ratios), with agreement within 20%. All significant features remained significant under conservative block bootstrap testing with Bonferroni correction.
Conclusions
We present a novel statistical framework for rigorous morphological analysis of 3D tumor spheroid time-series. By addressing temporal autocorrelation—a commonly overlooked source of statistical invalidity—the framework enables reliable hypothesis testing for longitudinal cancer imaging studies. This methodology has immediate applications in anti-cancer drug screening, where statistically valid phenotypic profiling is essential.
