Deep learning–enabled three-dimensional intelligent quantitative analysis for organoids (<italic>invited</italic>)

ObjectiveOrganoids are three-dimensional (3D) multicellular structures that recapitulate essential biological and functional characteristics of human tissues, serving as powerful in vitro models for drug screening and disease research. However, conventional bright-field microscopy cannot capture true 3D morphological dynamics during drug response. Fluorescence imaging requires exogenous labeling that may perturb the native microenvironment, while biochemical assays are destructive and preclude longitudinal observation of the same sample. These limitations hinder accurate, continuous, and high-throughput evaluation of organoid growth and therapeutic response. To address these challenges, this study aims to develop a label-free, scanning-based three-dimensional imaging and quantitative analysis system capable of long-term, in situ monitoring of organoid morphological evolution under drug treatment.MethodsA compact scanning optical acquisition system was constructed to operate directly inside a standard cell culture incubator, ensuring stable temperature, humidity, and CO2 conditions throughout long-term experiments. The optical architecture integrates a programmable illumination module and a motorized scanning stage to acquire multi-angle projection images across a large field of view. Controlled by custom-developed software, the scanning module sequentially captures image stacks under predefined angular configurations, generating raw datasets suitable for volumetric reconstruction. Three-dimensional structural information was reconstructed using a computational imaging pipeline based on optical diffraction tomography principles. Phase retrieval and inverse scattering algorithms were implemented to recover the refractive index distribution of organoids from multi-angle intensity measurements. To improve reconstruction stability, noise suppression and spatial frequency filtering were incorporated into the preprocessing stage. For quantitative analysis, a deep learning–assisted 3D segmentation framework was established. A convolutional neural network with an encoder–decoder architecture was trained on manually annotated volumetric datasets to accurately delineate organoid boundaries within reconstructed refractive index maps. Data augmentation strategies, including rotation, flipping, and intensity perturbation, were applied to enhance model generalization. The trained model enabled automated voxel-level segmentation of multiple organoids within a single volumetric dataset. To achieve longitudinal tracking, a temporal matching strategy combining spatial proximity and morphological similarity metrics was implemented. Segmented organoids at adjacent time points were associated based on centroid distance and volumetric consistency, allowing continuous identification of individual organoids over time. Quantitative morphological parameters—including volume, surface area, and sphericity—were calculated from the segmented 3D masks at each time point. Statistical analysis was conducted to evaluate temporal trends and inter-organoid variability under different drug treatment conditions.Results and DiscussionsThe proposed system achieved continuous 3D imaging of colorectal cancer organoids over several days while maintaining stable illumination, focus, and environmental conditions. In drug-free control groups, organoids exhibited progressive increases in volume and surface area, indicating sustained growth. In contrast, drug-treated groups demonstrated concentration-dependent growth inhibition and, in some cases, structural deformation or collapse, reflecting differential drug sensitivity. For representative organoids (n=5), volumetric measurements showed consistent increases over time under control conditions, accompanied by measurable inter-organoid variability, highlighting intrinsic biological heterogeneity. Statistical analysis across the full population confirmed a general upward trend in mean volume and surface area during culture, whereas sphericity remained relatively stable with minor fluctuations, suggesting preservation of overall structural integrity despite size expansion. Compared with conventional two-dimensional bright-field imaging and endpoint biochemical assays, the present system provides true volumetric information, improved temporal resolution, and non-destructive longitudinal tracking of individual organoids. The integration of deep learning–assisted segmentation enhances quantitative robustness and reduces manual intervention, enabling reliable high-throughput analysis. These results demonstrate the platform’s capability to capture subtle morphological dynamics and accurately evaluate drug-induced structural changes.ConclusionsA label-free, high-throughput, and long-term three-dimensional imaging and quantitative analysis platform for colorectal cancer organoids has been developed and experimentally validated. The system enables accurate extraction of key morphological parameters and supports continuous monitoring of structural evolution during drug treatment. The platform demonstrates strong stability, reconstruction fidelity, and analytical reliability under incubator conditions. Nevertheless, further optimization is required to improve scanning speed, enhance robustness under challenging imaging scenarios, and expand the diversity of quantitative morphological descriptors. Future work will focus on adapting the optical design for multi-well plate compatibility, enriching neural network training datasets to improve generalization across organoid types, and integrating the system with automated culture modules and intelligent analysis software. The proposed system shows significant potential for large-scale drug screening applications and personalized medicine research, providing a stable and efficient technological framework for high-throughput, in situ 3D organoid analysis.