Cell Detection and Segmentation Using Correlation Clustering

Chong Zhang, Julian Yarkony, Fred A. HamprechtMICCAI 2014 · DOI

TL;DR

This paper introduces a novel method for cell detection and segmentation in microscopy images, particularly effective for transparent cells. By combining a learned boundary classifier with a constrained correlation clustering approach, they achieve state-of-the-art performance on challenging datasets, improving robustness in crowded cell environments.

What Problem Does This Paper Address?

Analyzing microscopy images is crucial for high-throughput biological experiments, but accurately detecting and segmenting cells, especially those that are nearly transparent or appear in dense clusters, remains a significant challenge. Existing methods often struggle with varying cell shapes, modalities, and incomplete boundary information. This paper aims to provide a robust and automated solution for these issues, enabling more precise cell-level quantification.

Methods

The proposed method consists of three main steps:

  1. Boundary Evidence Extraction: A classifier (specifically, a random forest trained using the ilastik toolkit) is trained to predict the probability of each pixel belonging to a cell boundary. This step requires minimal manual labeling (<1% of pixels).
  2. Superpixel Computation: The predicted boundary probabilities are used to generate initial regions called superpixels. The watershed transform is employed, with a smoothed probability map to avoid tiny, non-informative superpixels and to separate background from potential cell regions. The smoothing is guided by the average cell length.
  3. Correlation Clustering Segmentation: A weighted region adjacency graph is constructed where superpixels are nodes and edges represent adjacency. Edge weights are derived from boundary probabilities, with negative weights encouraging separation and positive weights encouraging merging. The core of the segmentation is a correlation clustering (or multicut) problem, which partitions the graph to minimize the total weight of cut edges. A key novelty is the introduction of a length constraint, formulated using a "Semi-PlanarCC" model. This constraint penalizes segmentations where cells exceed a certain maximum length, which is crucial for crowded or partially occluded cells. This constraint is approximated using non-planar correlation clustering.

Datasets

The method was evaluated on three datasets:

  • Dataset-a: Bright field images of Diploid yeast cells (1K×1K pixels, cell size 40-100 pixels).
  • Dataset-b: Bright field images of Fission yeast cells (1K×1K pixels, cell size 30-140 pixels), similar to those used in [9].
  • Dataset-c: Phase contrast images of cervical cancer cells (HeLa cell line) (400×400 pixels, cell size 10-40 pixels), similar to those used in [3].

These datasets represent different cell types, imaging modalities (bright field, phase contrast), and cell densities, allowing for a comprehensive evaluation.

Key Insights and Ideas

  • Learning Boundary Evidence is Key: Training a classifier to predict boundary probabilities, even with sparse labels, provides a flexible and powerful input for subsequent segmentation.
  • Correlation Clustering for Segmentation: This graph-based partitioning method is well-suited for image segmentation, especially when dealing with complex relationships between regions.
  • The Power of Length Constraints: Incorporating a length constraint (via Semi-PlanarCC) significantly improves performance, particularly in scenarios with overlapping cells and missing boundaries. This constraint leverages prior biological knowledge about cell sizes.
  • Handling Transparent Cells: The method is specifically designed to work well with cells that are almost transparent, relying on subtle intensity changes at boundaries.
  • Robustness to Different Modalities: The approach demonstrates applicability across different microscopy techniques and cell morphologies.

Key Results

  • The proposed method, particularly the Semi-PlanarCC variant with the length constraint, achieves state-of-the-art performance on the evaluated datasets.
  • In detection tasks, the method shows improved recall compared to standard PlanarCC, especially in challenging scenarios.
  • Segmentation accuracy, measured by area overlap metrics (M1 and M2), is competitive with or superior to other methods like CellDetect, CellX, and CellStat, especially for datasets with complex cell arrangements.
  • The method is efficient enough for high-throughput studies, requiring minimal user interaction beyond initial classifier training.
  • Qualitative results show that out-of-focus cells are effectively excluded from segmentation.

Limitations and Future Work

  • Post-processing: The authors acknowledge that deterministic post-processing steps (e.g., using shape or texture features) could further refine segmentations and reject segments with implausible characteristics.
  • Parameter Tuning: While the method is largely automatic, some parameters (like the bias term 't' and the length constraint 'd') require tuning. The paper suggests that Semi-PlanarCC is less sensitive to 't' than PlanarCC.
  • Future Work: The paper mentions that deterministic post-processing steps should be applied to further reject segments with implausible characteristics. They also focus on evaluating the performance of their technique, implying future work could explore more advanced post-processing or integration with other biological feature analysis.

Summary updated

BibTeX
@inproceedings{zhang2014cell,
  author = {Chong Zhang and Julian Yarkony and Fred A. Hamprecht},
  title = {{Cell Detection and Segmentation Using Correlation Clustering}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2014},
  doi = {10.1007/978-3-319-10404-1\_2},
}
Figure