Automated Live-Cell Imaging Analysis (ACIA)

• ACIA is an end-to-end framework that converts raw time-lapse images into per-cell quantitative data using automated segmentation, tracking, and adaptive control.
• It integrates diverse imaging modalities—including fluorescence, phase-contrast, and label-free techniques—across complex platforms like microfluidics and high-throughput systems.
• Advances in deep learning and design-aware processing enhance real-time performance and accuracy while addressing challenges like phototoxicity and sample health.

Automated live-Cell Imaging Analysis (ACIA) denotes the conversion of raw time-lapse microscope images into quantitative, per-cell readouts with minimal human intervention and, increasingly, in near real time. In microfluidic live-cell imaging this is especially demanding because thousands of spatially structured regions of interest can be embedded in complex device architectures, but the same term also encompasses label-free phase-contrast workflows, fluorescence time series, quantitative phase imaging, droplet microfluidics, and event-driven microscopy. Across these settings, ACIA is best understood as an end-to-end chain that joins sample handling, image acquisition, computational reconstruction, segmentation, tracking, quantification, visualization, and, in some systems, closed-loop experimental control (Seiffarth et al., 16 Jun 2026, Wei et al., 2023).

1. Conceptual scope and historical development

ACIA emerged from a convergence of automated microscopy, quantitative image analysis, and domain-specific workflow engineering. Early integrated systems already combined segmentation, tracking, and corrective visualization at scale. LEVER 3-D, for example, handled multichannel 5-D $(x,y,z,t,\text{channel})$ confocal fluorescence microscopy, automatically segmented, tracked, and lineaged stem-cell image sequences, and then allowed the user to edit results in a stereoscopic 3-D window while viewing the lineage tree in 2-D; its validation interface was designed so that each data set could be corrected to 100% accuracy (Wait et al., 2014). In parallel, specialized live-cell modules such as MitosisAnalyser used a circular Hough transform plus variational tracking to detect and follow mitotic cells in phase-contrast sequences, yielding mitosis duration, cell fates, and morphology statistics without requiring fluorescent mitotic markers (Grah et al., 2016).

A second strand of development centered on particle-centric live-cell analysis. Cellulyzer, implemented primarily as an ImageJ plugin, automated tracking and comparative analysis for microtubule plus-end proteins, centrosome pairs, and diffusive trajectories, and coupled these analyses to interactive simulation so that imaging parameters could be chosen using feedback from the final quantitative readouts (Jafarpour et al., 2017). A related tendency was to push automated analysis into more constrained or translational settings, such as microfluidic blood-cell detection via Faster R-CNN for point-of-care devices, where per-frame detection and counting serve as the core of a broader live-cell analysis stack (Xia et al., 2019).

Recent work broadens the term still further. Reviews of computation in live-cell microscopy emphasize that denoising, super-resolution, spectral unmixing, adaptive optics, and event-driven acquisition now alter the quality and structure of the data before segmentation and tracking even begin (Shroff et al., 2024). In the same period, photodamage-aware perspectives have argued that live-cell automation should not only maximize image quality, but also explicitly account for sample health, because none of the existing AI solutions analyses or estimates it directly (Gómez-de-Mariscal et al., 2023). This suggests that ACIA is no longer just automated image analysis in the narrow sense, but an increasingly experiment-aware computational framework.

2. Experimental substrates and imaging architectures

A defining feature of ACIA is that the computational pipeline is often inseparable from the physical substrate. In droplet microfluidics, Auto-ICell is an end-to-end automated live-cell imaging and analysis system built around a stereolithography-printed 3D microfluidic chip with glass capillaries, syringe-pump control of oil and aqueous phases, collection into 96-well plates, and alternating bright-field and fluorescence microscopy. Its co-flow droplet generator produces water-in-oil droplets with diameters from 70 $\mu$m to 240 $\mu$m at approximately 1,500 droplets per minute, so each droplet can function as an isolated microreactor for single-cell morphology and apoptosis analysis (Wei et al., 2023).

In high-throughput microfluidic cultivation chips, the substrate itself becomes part of the computational prior. DART is a design-aware and real-time capable paradigm that aligns a chip’s CAD blueprint with the physical chip by embedding fiducial markers and then using deep-learning-based marker detection. On the Swiss Army Knife chip, which contains 1164 regions of interest across eight different geometries, DART localizes all regions of interest in five minutes, removes microfluidic structures from raw microscopy images in 40 ms, and performs fully automated image analysis, including cell segmentation, in under 1.1 s per image (Seiffarth et al., 16 Jun 2026). The crucial architectural point is that localization and masking are no longer inferred from image content alone; they are anchored in the known device design.

Another branch of ACIA is built around label-free optics. PICS combines quantitative phase imaging with AI so that co-localized fluorescence can be replaced by virtual stains predicted from SLIM or GLIM phase images. Because inference is built into the acquisition software and runs in real time, the same platform can monitor nuclei and cytoplasm over many days without loss of viability, and can also quantify nuclear dry mass inside spheroids where multiple scattering would otherwise limit conventional imaging (Kandel et al., 2020). This label-free strategy is complemented by resource-constrained implementations such as DataSet Tracker, a Unity/OpenCV software suite designed for real-time analysis on computers, smartphones, and smart glasses hardware and suitable for microscopes without internet connectivity (Matov, 2024).

These architectures indicate that ACIA is not tied to a single modality. It includes bright-field and multi-channel fluorescence in droplets, phase-contrast time-lapse analysis, 5-D confocal microscopy, quantitative phase imaging, and low-cost edge computing. The unifying principle is that acquisition hardware and analysis software are co-designed so that segmentation, tracking, and quantification become operationally routine.

3. Core computational operations

The computational core of ACIA is organized around a small set of recurring operations: segmentation, tracking, region masking, feature extraction, and classification. In classical fluorescence workflows, watershed-based segmentation remains a central primitive. A GUI-oriented watershed algorithm for fluorescence microscopy combined adaptive histogram equalization, background subtraction, smoothing, multilevel Otsu background detection, and post-segmentation area and signal constraints so that touching cells could be segmented and then classified by per-object channel intensities (Bartell et al., 2017). In phase-contrast microscopy, more specialized formulations were required: MitosisAnalyser initialized contours with the circular Hough transform and propagated them with a tailored level-set variational method to detect mitotic entry, follow mitosis backward and forward in time, and assign outcomes such as division, no division, or cell death (Grah et al., 2016).

Deep learning now dominates many segmentation stages, but ACIA systems differ in how much contextual structure they expose to the model. Auto-ICell uses a Python pipeline in which Cellpose-based deep learning generates single-cell masks from fluorescence images, after which area, perimeter, circularity, and apoptosis-related intensities are computed automatically (Wei et al., 2023). DART deliberately moves chip-geometry handling out of the segmentation model: design-aware masking removes channels, walls, pillars, and other structures before Cellpose-SAM is applied, so the model sees a simplified single-region image rather than raw microfluidic clutter (Seiffarth et al., 16 Jun 2026). PICS pushes computation even earlier in the chain by using a modified U-Net to infer virtual fluorescence channels from quantitative phase maps; those predicted channels then function as semantic masks for compartment-specific measurements (Kandel et al., 2020).

Tracking has undergone a similar shift from hand-tuned optimization to learned association. Trackastra uses a transformer architecture that operates on detections within a temporal window and directly learns pairwise associations of cells, including dividing objects. Its lineage-aware normalization is encoded by the parental softmax

$$\tilde{A}_{ij}=\frac{\exp(\hat{A}_{ij})}{1+\sum_{i' \in \mathcal{P}_j}\exp(\hat{A}_{i'j})},$$

which enforces at most one parent per child while still allowing a parent to produce two daughters (Peoples et al., 2024). This removes much of the dataset-specific hyperparameter tuning that characterized earlier tracking-by-detection pipelines.

Not all high-performing ACIA systems are training-based. A 2025 low-cost GUI for unstained cell-culture analysis explicitly avoided manually annotated training data and a training phase, instead combining neighborhood-based statistical filtering, fuzzy intensity transformation, local statistics, Moran’s $I$, variogram-based measures, semantic segmentation, contour-based instance segmentation, and automated report generation on standard CPU hardware (Das et al., 14 Sep 2025). This is a useful reminder that ACIA includes both learned and non-learned pipelines, provided they deliver reproducible, quantitative, low-intervention analysis.

4. Quantification, metrics, and real-time performance

ACIA is defined as much by its output variables as by its segmentation masks. In Auto-ICell, morphological quantification is based on area, perimeter, and circularity,

$$\text{Circularity}=\frac{4\pi \times \text{Area}}{\text{Perimeter}^2},$$

and apoptosis is quantified from Calcein-AM and PI channels using per-cell fluorescence ratios, population live/dead fractions, and time-dependent intensity traces. The same platform reports real-time display within less than 2 seconds per frame, Cellpose-based cell detection in about 0.194 s per image, and end-to-end capture-to-display latency below 1.772 s (Wei et al., 2023).

DART emphasizes throughput-independent localization and quantitative growth analysis. After design-aware masking and segmentation, it computes Total Single-Cell Area (TSCA) per region of interest and fits logistic growth curves to TSCA time series. In the reported experiment, 1739 images were processed in 31 minutes, more than 500,000 individual cells were segmented, and the total pipeline time was $1.067 \pm 0.697$ s per image, corresponding to a real-time factor of approximately 17.4 relative to acquisition (Seiffarth et al., 16 Jun 2026). The metric design here is notable: the primary quantitative observable is not merely a cell count, but a per-region time series that can support downstream growth modeling.

EAP4EMSIG makes the latency–accuracy trade-off explicit. In its 2024 segmentation benchmark for bacterial microcolonies, Omnipose achieved a Panoptic Quality of 0.9336, while Contour Proposal Network achieved the fastest inference time of 185 ms with a Panoptic Quality of 0.8575 (Friederich et al., 2024). The 2025 extension added a deep-learning autofocus module with a Mean Absolute Error of 0.0226 $\mu$m and inference times below 50 ms, and expanded segmentation benchmarking to eleven methods, where Cellpose 3 reached a Panoptic Quality of 93.58% and a distance-based method reached 121 ms with a Panoptic Quality of 93.02% (Friederich et al., 30 Mar 2025). These results show that ACIA performance is increasingly characterized by a joint metric space comprising segmentation quality, inference time, autofocus accuracy, and compatibility with feedback control.

Other systems highlight additional performance envelopes. DataSet Tracker reported 5 fps in a desktop demo and 19 fps in a glial nuclear marker tracking dataset while producing real-time optical-flow vectors, speed histograms, and direction histograms on resource-constrained platforms (Matov, 2024). PICS reported 65 ms inference per virtual stain, with the computational chain shorter than the acquisition chain, and a nuclear dry-mass fraction discrepancy of about 4% between PICS-based and DAPI-based masks in spheroids (Kandel et al., 2020). These figures indicate that real-time ACIA can now operate across very different latency regimes, from sub-50 ms focus control to second-scale whole-image segmentation and day-scale longitudinal dry-mass monitoring.

5. Event-driven control, adaptive imaging, and photodamage awareness

A major contemporary shift is that ACIA is moving from post hoc analysis toward active control of the microscope. EAP4EMSIG formalizes this as a cyclic experiment-automation loop with microscope control, image acquisition, real-time image processing, OMERO storage, semi-automatic annotation, simulation, real-time data analysis and event detection, and a real-time experiment planner that selects the next chambers to image and sends instructions back to microscope control (Friederich et al., 2024). Technical events such as focus loss and chamber defects, and biological events such as changes in growth rate, cell death, and chambers reaching capacity, are detected from segmentation-derived features and can trigger altered acquisition schedules (Friederich et al., 30 Mar 2025).

DART was designed with the same destination in mind. Because the CAD blueprint is aligned with the physical chip and each region of interest can be localized independently of throughput, design-aware masking and segmentation produce quantitative readouts quickly enough to support closed-loop and outcome-driven smart microscopy (Seiffarth et al., 16 Jun 2026). Inference is therefore no longer downstream of the experiment; it becomes part of the experiment’s control logic.

The strongest motivation for this shift is phototoxicity and photon-budget management. The photodamage-aware perspective on live imaging argues that AI can increase information content from low-dose data, support adaptive and event-driven acquisition, and ultimately feed real-time decisions that balance sample health and image quality. At the same time, it states explicitly that none of the existing solutions analyses or estimates sample health directly, and it calls for robust photodamage reporters that provide quantitative assessments of sample health without requiring additional fluorescence channels (Gómez-de-Mariscal et al., 2023). This establishes an important conceptual distinction within ACIA: high image quality is not equivalent to biologically safe imaging.

Label-free and low-dose approaches address this distinction from different directions. PICS removes most fluorescence exposure from long-term monitoring by transferring specificity from terminal fluorescence labels to quantitative phase inference (Kandel et al., 2020). DataSet Tracker, by contrast, uses optical flow and spot detection on standard microscope feeds, including devices without internet connectivity, to provide immediate overlays and quantitative feedback that can guide experimental actions in real time (Matov, 2024). Together these systems suggest that ACIA is evolving toward an overview of analysis, control, and sample-preserving acquisition.

6. Applications, limitations, and open directions

The application space described for ACIA is broad. Auto-ICell highlights cell culture, biochemical microreactors, drug carriers, cell-based assays, synthetic biology, and point-of-care diagnostics, while demonstrating long-term 3D culture and apoptosis monitoring of MDA-MB-231 breast cancer cells in droplets (Wei et al., 2023). DART targets microfluidic cultivation chips with hundreds to thousands of microbial populations and positions itself as a route toward closed-loop and outcome-driven smart microscopy in high-throughput microfluidics (Seiffarth et al., 16 Jun 2026). PICS extends automated analysis to long-term, label-free monitoring of nuclei and cytoplasm and to dry-mass measurements inside spheroids (Kandel et al., 2020). DataSet Tracker emphasizes cytoskeletal dynamics, vesicle trafficking, organoids, circulating tumor cells, and drug-response phenotyping on modest hardware (Matov, 2024). Historical tools such as LEVER 3-D, Cellulyzer, and MitosisAnalyser illustrate corresponding uses in stem-cell lineage analysis, microtubule plus-end dynamics, centrosome pairing, diffusion processes, and phase-contrast mitosis analysis [(Wait et al., 2014); (Jafarpour et al., 2017); (Grah et al., 2016)].

The literature also records explicit limits. Auto-ICell demonstrations focus on MDA-MB-231 breast cancer cells, and the fluorescence implementation is mainly centered on two dyes, Calcein-AM and PI, plus generic fluorescence for morphology (Wei et al., 2023). DART depends on CAD-defined geometries and embedded fiducial markers, so its strongest advantages arise where the physical device is digitally specified in advance (Seiffarth et al., 16 Jun 2026). Trackastra reduces tuning at the linking stage, but still depends on the quality of upstream detections because it operates on segmentations or detections rather than repairing them directly (Peoples et al., 2024).

A more pointed controversy concerns foundation models and generative enhancement. In the 2024 EAP4EMSIG benchmark, Segment Anything was unsuitable for real-time segmentation of dense bacterial microcolonies, and in the 2025 extension all six Deep Learning Foundation Models were unsuitable for real-time segmentation (Friederich et al., 2024, Friederich et al., 30 Mar 2025). The phototoxicity review makes the complementary point that deep models can hallucinate, so virtual structures or intensities may not correspond to real biology; it therefore calls for rigorous validation with orthogonal techniques and explicit quality flags in AI-generated images (Gómez-de-Mariscal et al., 2023). These findings do not negate the value of AI in ACIA, but they do reject the view that large general-purpose models can simply be substituted for domain-specific validation.

Future directions are correspondingly hybrid. Auto-ICell suggests higher-resolution 3D printers, multiple-emulsion or all-aqueous emulsion strategies, stain-less image analysis, virtual histological staining, more advanced models, additional biomarkers, and closed-loop experiments in which real-time analysis triggers interventions (Wei et al., 2023). DART generalizes the idea of design-aware analysis to other structured devices and structured imaging platforms (Seiffarth et al., 16 Jun 2026). Trackastra points toward joint segmentation-and-tracking and stronger use of non-adjacent temporal associations (Peoples et al., 2024). The photodamage-aware literature argues for label-free phototoxicity estimation, transfer learning and domain adaptation, and systematic validation pipelines (Gómez-de-Mariscal et al., 2023). A plausible implication is that mature ACIA will not be a single algorithmic category, but a layered infrastructure in which acquisition physics, device geometry, reconstruction, segmentation, tracking, and experiment control are all explicitly coupled.