Autofocusing is an essential part of fully automated image cytometry systems. Defocusing of a microscope mainly reduces the middle frequencies; just below N U . The signal power after applying a band pass filter centered around NA/L gives the best focus function. Which digital filter selects this band depends on the sampling density. When sampling at the Nyquist rate a simple gradient filter { 1 D-11 performs well. Higher order derivatives obtained by repeating this filter makes the pass band smaller and therefore sharpens the peak of the corresponding focus function. Undersampling by a factor of two, e.g. through binning on the CCD array, produces similar results using a { 1, - I ] filter. This selects exactly the same part of the spectrum, has an improved S N R and uses only 25% of the data points. Using this filter we designed a three phase autofocus algorithm. The first phase quickly finds the region near focus. This phase becomes redundant when scanning a slide since it is never far from focus. The second phase roughly locates the peak and the third phase applies a quadratic fit through focus values around the estimated peak position,
The quadratic fit uses only 3 to 5 data points depending on the SNR. The focus error remains smaller than the reproducibility of the z-axis (50 nm). The procedure works well even for light levels down to 200 photons per pixel before binning. Scanning PI labeled nuclei using a Photomeuics KAF1400 camera and a Nikon 20/0.75 objective, five images (100 x 100) with 0.2 s integration time where needed to focus. This research was supported by NWGMedigon and Imagenetics.