Automated Cell Detection

Introduction

Unbiased stereological methods can be labor-intensive, even using computer-controlled microscope equipment. In an ongoing effort to increase the efficiency of data analysis, we are investigating a new methodology for automated cell recognition. Previously we demonstrated an improvement in data collection rate by integrating automated cell detection with the unbiased 3D counting frame of the Optical Fractionator in the stereology program, Stereo Investigator®. Adaptive local thresholding was used to detect cells based on minimum diameter. However, tightly packed cells were inaccurately detected as large single cells. To improve cell detection accuracy, we present an approach with non-overlap constraints for handling aggregated cells, where cells are modeled as deformable 3D ellipsoids. Results of the automated detection may be inspected and edited by a human to ensure accuracy. The goal of automated stereology software is to improve the data-collection rate for stereological analyses, while providing more detailed morphological analyses.

Potential Advantages of Automated Detection

  • Fast
  • Unbiased
  • Adaptable to specific applications
  • Produce shape descriptions (e.g., volume, sphericity)
  • Measures precise position in tissue for cross-channel colocalization analyses

Automated Cell Detection

The algorithm uses a model-based framework with user-defined parameters and input. The three-dimensional (3D) object detector uses:

  • Adaptive local thresholding to balance the trade-off between high sensitivity for faint cell detection and high false-positive rates due to noise in the image.
  • Marker Initiated Watershed Segmentation to eliminate other smaller objects in the image such as dendrites and spines. It provides initial splitting of touching cells.
  • Model based cluster division to model the cells based on shape and non-overlap constraints to efficiently divide clustered cells.
  • Object Classification to validate cells and exclude irrelevant background artifacts based on geometry and intensity.

Colocalization

Automatically detected cells, represented in 3D, enable cross-channel colocalization analyses. Types of colocalization include volumetric overlap, centroid-to-centroid distance or Pearson’s colocalization coefficient.

Conclusions

We have developed an improved automated cell detection method suitable for use with the Optical Fractionator method for estimation of total cell number. Streamlined in a workflow, the method relies on computer-controlled image acquisition in a systematic-uniform-random series followed by automatic detection of cells within the stereological probe together with visual 3D inspection. It ensures counting rules are consistently applied while reducing the manual labor required with current methods. Precise marker placement by the algorithm ensures that the counting rules are objectively and consistently applied throughout the experiment. However, manual stereology requires a subjective decision about whether a cell is within the counting frame. Therefore, there will be some incongruencies between human and machine.

Automated Cell Detection will be available as an extension module in 2010.