A statistically rigorous multi-scale texture analysis framework for 3D spheroid characterization: temporal autocorrelation correction and molecular validation

Three-dimensional tumor spheroids represent dynamic biomodeling systems exhibiting emergent collective behaviors-spontaneous reorganization, migration dynamics, and epithelial-mesenchymal transition observable through label-free time-lapse microscopy. However, optical opacity from light scattering limits fluorescence imaging depth, preventing single-cell resolution deep within spheroid volumes and necessitating ensemble-level morphological analysis rather than exhaustive single-cell profiling, while destructive molecular endpoint assays provide only discrete temporal snapshots, missing continuous dynamic changes defining biological processes. We present a validated computational framework for statistically rigorous ensemble morphological profiling, integrating multi-scale texture analysis (gray-level co-occurrence matrices, wavelet decomposition, Gabor filtering; 37 features spanning orientations and scales), global standardization eliminating scale artifacts while preserving biological signal, and autocorrelation-informed block bootstrap resampling addressing temporal dependence. Proof-of-concept analysis using representative A549 and H1299 lung cancer spheroids (n = 1 per condition, 48 hourly observations) demonstrates three framework capabilities. First, global standardization normalized features spanning 24 orders of magnitude in variance while preserving biological discrimination (16.8-fold Fisher score difference between cell lines). Second, temporal autocorrelation analysis revealed 4-hour median decorrelation lags, addressed through 5-h block bootstrap resampling enabling valid statistical inference. Third, external validation using Cancer Dependency Map RNA-sequencing (1699 cell lines) demonstrated that texture discrimination (16.8-fold) corresponded quantitatively to independent molecular differences (14.6-fold VIM/CDH1 ratio), achieving concordance within 15%. This framework enables statistically valid, biologically grounded morphological profiling for continuous monitoring applications including drug screening, organoid development tracking, and biomanufacturing quality control where ensemble dynamics complement discrete molecular measurements.