Introduction
In cancer research and drug discovery, understanding the precise impact of genetic or chemical perturbations on cell division (mitosis) is crucial. While modern high-throughput microscopy can generate vast amounts of data, manual analysis remains a significant bottleneck. Traditional automated tools often focus on population-level averages, which can obscure subtle but critical individual cell behaviors. Furthermore, many existing methods are designed only for "normal" cell cycles, failing to accurately categorize the abnormal phenotypes—such as chromosome misalignment or nuclear fragmentation—that are often the primary targets of drug testing.
To address these challenges, this paper presents an automated pipeline designed to track and analyze individual cells throughout their entire division cycle. By integrating machine learning with a biological "logic engine," the system identifies both standard mitotic stages and five distinct abnormal phenotypes induced by external interference.
Methods
The proposed computational pipeline consists of four main stages: segmentation and tracking, feature extraction, classification, and temporal logic filtering.
Segmentation and Tracking
The system utilizes a region-adaptive thresholding technique to isolate cell nuclei. A specialized post-processing step was developed to handle "messy" cells by re-attaching small, detached chromosomes to their parent nuclei. Tracking is performed in two phases: first, cells are matched frame-to-frame; second, "mitosis events" (where a single cell divides into two) are detected using a likelihood measure based on changes in size and brightness.
Feature Extraction and Classification
For each identified nucleus, the system extracts 376 features, including Haralick texture features, Zernike shape moments, and dynamic features that capture temporal changes in appearance. These features serve as input for a Support Vector Machine (SVM) trained to classify cells into one of 12 categories: seven normal mitotic phases (from interphase to telophase) and five abnormal phenotypes, such as abnormal anaphase or cell death.
Temporal Logic (The Finite State Machine)
To ensure biological plausibility, the pipeline incorporates a Finite State Machine (FSM). This "logic engine" prevents impossible transitions (e.g., a cell reverting from telophase to prophase) and acts as a filter to correct transient classification errors. This temporal context allows for precise measurement of the duration of each mitotic stage.
Results
The system was validated using 48 3D image sequences of HeLa cells expressing histone EGFP. The dataset included control groups, cells treated with varying doses of Nocodazole, and cells with ch-TOG gene knockdown via siRNA.
Robust Performance
The pipeline demonstrated high technical reliability, achieving 98.1% segmentation accuracy and 99.8% tracking accuracy. The SVM-based classification maintained approximately 94% accuracy across both drug-treated and gene-knockdown datasets, proving its robustness in the presence of significant morphological abnormalities.
Biological Validation of Temporal Phenotypes
The tool successfully quantified "temporal phenotypes," revealing that both chemical and genetic perturbations primarily manifest as delays in specific mitotic stages.
As shown in the figure above, the system identified that high-dose Nocodazole treatment and ch-TOG knockdown both caused a massive delay in prometaphase—the stage where chromosomes prepare to align. Specifically, high-dose Nocodazole resulted in a highly significant prolongation ($p = 4.2 \cdot 10^{-12}$), while ch-TOG knockdown showed an even stronger effect ($p = 2.3 \cdot 10^{-12}$). These results suggest that phase duration is a more sensitive indicator of drug effects than morphology alone.
Discussion
While the system provides a powerful tool for phenotypic screening, some limitations remain. The current study utilized 3D sequences with limited Z-resolution (three slices per time step) to maintain high-throughput efficiency. Future work could explore higher resolution data for more detailed depth information. Additionally, as the number of modeled phenotypes increases, the complexity of the FSM grows; scaling to more complex behaviors might necessitate probabilistic models like Hidden Markov Models. Finally, while validated on HeLa cells, further testing is required to determine how well the extracted features generalize to other cell types with different nuclear morphologies.
In conclusion, by treating abnormalities as valuable data rather than noise and leveraging temporal context, this automated pipeline offers a robust method for detailed mitotic analysis in drug discovery and basic cell biology.
Summary updated