AUTOMATED CELL RECOGNITION AND COUNTING BASED ON A COMBINATION OF ARTIFICIAL NEURAL NETWORKS AND STANDARD IMAGE ANALYSIS METHODS
Per Jesper Sjöström1 and Lars Ulrik Wahlberg2
1) Brandeis University, Biology Department,
415 South Street, Waltham MA 02454-9110, firstname.lastname@example.org
2) CytoTherapeutics Inc., 701 George Washington Highway, Lincoln RI 02865, email@example.com
Standard image analysis tools (i.e. methods
that rely on textures, boundary contours, histogram thresholding,
gray-scale intensity peak search, etc.) did not work well with
our noisy histological preparations. To avoid time-consuming manual
cell counts, an automated cell counter was constructed using a
combination of artificial neural networks (ANN) and standard imaging
methods. An ANN was applied directly on digitized microscopy fields
without pre-ANN feature extraction. A three-layer feed-forward
network was trained using the error back-propagation algorithm
to recognize cells in digitized 48-by-48-pixel images. The 1830
manually selected training samples consisted of the two classes
cells and non-cells. The former were images of well-centered
live cells, whereas the latter were images of debris, uncentered
cells, background, etc. The degrees of freedom of the ANN was
minimized by extensive weight sharing in the input layer, and
by estimating the optimal number of neurons in both hidden layers.
The trained ANN was used to filter out non-cell objects
from digitized microscopy fields, and standard imaging tools subsequently
produced the cell count. The trained system was validated by comparison
with blinded human counts. The ANN training software was implemented
in Pascal on a Power Macintosh 7300/180, and the cell counting
system was based on a modified version of NIH Image. The correlation
index at 100x magnification neared person-to-person variability.
The system was approximately six times faster than a human. In
conclusion, ANN-based cell counting using a standard desktop computer
is feasible. Consistent histology was crucial to system performance.
User-friendliness would profit from increased computer power.
Two benefits of the system were speed of analysis and consistency
of cell counts.