Abstract
In the effort to aid cytologic diagnostics by establishing automatic single cell screening using high throughput digital holographic microscopy for clinical studies, thousands of images and millions of cells are captured. The bottleneck lies in an automatic, fast, and unsupervised segmentation technique that does not limit the types of cells which might occur. We propose an unsupervised multistage method that segments correctly without confusing noise or reflections with cells and without missing cells, including the detection of relevant inner structures like the cell nucleus in unstained cells. To make the information interpretable for cytopathologists, we introduce new cytoplasmic and nuclear features exploiting quantitative phase information. We show that the segmentation provides consistently good results over many experiments on patient samples in a reasonable per-cell analysis time.
Introduction
The traditional PAP test for cervical cancer screening is a labor-intensive manual process prone to human fatigue. While Quantitative Phase Microscopy (QPM) allows for high-resolution, label-free imaging of cells, it creates a massive data bottleneck. Existing segmentation tools often require supervised training data or struggle with optical artifacts. Most importantly, identifying the nucleus in an unstained cell is difficult without chemical dyes. This paper presents an unsupervised, multistage algorithm designed to segment millions of cells and their nuclei from QPM data efficiently.
Methods
The proposed pipeline consists of four unsupervised stages:
- Statistical Analysis: The algorithm analyzes phase values across the measurement set to automatically determine background and noise levels.
- Threshold Stage: A coarse global threshold is applied to identify potential cell candidates.
- Plausibility Checks: Candidates are filtered based on area (radius > 3 µm), hierarchy (removing nested contours), and gradient verification to ensure physical boundaries.
- Internal Structure Detection: A second segmentation pass within the cell boundary uses relative thresholds to pinpoint the nucleus based on its higher optical density.
Results
The algorithm was tested on a dataset of 100 patient samples, encompassing approximately 23 million individual cells. The analysis is remarkably fast, averaging 0.113 seconds per cell.
Key performance metrics include:
- Missed Cells (False Negatives): 1.85%, a critical metric for cancer screening.
- False Positives: 5.94%, largely attributed to debris or slime.
- Nuclear Detection: Successful in the majority of cases, though focus issues hindered segmentation in 13.27% of cells.
The system extracts 68 features per cell, including novel optical density metrics that mimic the morphological cues used by cytopathologists, such as the nuclear-cytoplasmic ratio.
Discussion and Conclusion
The unsupervised nature of this method is essential for clinical scalability, as it adapts to the statistics of specific measurements without requiring manual ground truth. While challenges remain regarding focus sensitivity and the exclusion of very small cells, the pipeline successfully translates raw phase data into interpretable diagnostic features. Future work will involve using these unsupervised results to bootstrap more complex deep-learning models and refining nuclear boundaries with active contour methods. This approach represents a significant step toward automated, high-throughput cytologic screening.
Summary updated