Phase contrast image segmentation by weak watershed transform assembly

We present here a method giving a robust segmentation for in vitro cells observed under standard phase-contrast microscopy. We tackle the problem using the watershed transform. Watershed transform is known for its ability to generate closed contours and its extreme sensitivity to image borders. One main drawback of this method is over-segmentation. In order to circumvent this, marked watershed based on the "modified gradient" method has been developed. However, the choice of the watershed mark locations is critical and their inadequacy may cause wrong results. Similarly to randomization and combination procedures used in the machine learning field, the present paper promotes the use of an assembly of marked watershed transforms, in order to increase the segmentation robustness. This results in the definition of candidate segmentations margins (expressed in terms of object border confidence) from which final segmentation can be chosen by means of thresholding.