WEKA for microscopical analysis of population composition: Training WEKA on multiple images that may or may not contain several out of multiple classes does not work

Hi,

I´m trying to use WEKA segmentation to measure population dynamics of different fungi in mixed cultures. I developed a dual staining procedure, that sufficiently discriminates different fungi based on their cell wall composition. I have a mixed culture of a yeast together with a filamentous fungus growing on a medium containing solid cellulose particles. Additionally, some samples cointain a third filamentous fungus. My goal is to study the population development by segmenting microscopical images and measure the projected area of the different classes. In my case that would be 5 classes:

Yeast, Fungus A, Fungus B, Cellulose, Background

I managed to train WEKA manually on one image that contains all 5 classes and when applying the classifier to different images containing all 5 classes, the results are acceptable.

Original RGB image:
CBP-TOM E-Extended Depth of Focus-01.czi - CBP-TOM E-Extended Depth of Focus-01.czi #01-1 RGB chan1 0-500 chan 2 0- 10000-1.tif (8.0 MB)

Segmented image
CBP-TOM E-Extended Depth of Focus-01.czi - CBP-TOM E-Extended Depth of Focus-01.czi #01-1 RGB chan1 0-500 chan 2 0- 10000-1.tif (2.7 MB)

However, on images where e.g. only yeast cell are visible, I get a lot of wrong classification. How can I further train the classifier on those images that contain only some but not all of the classes? I tried to open an image with only yeast in WEKA and loaded the trained classifier. Then I tried to train it further on this yeast image and saved the improved classifier afterwards. However, when I then used that classifier to segment an image with all 5 classes, I got an image that was only segmented in two classes (background and everything else). What did I do wrong so that the classifier “forgot” all the other classes during training?

Segmented image only yeast:
CBP-TOM E-Extended Depth of Focus-01.czi - CBP-TOM E-Extended Depth of Focus-01.czi #01-1 RGB chan1 0-500 chan 2 0- 10000-6.tif (2.7 MB)

Instead of training the classifier manually on images containing the wholes mixture, I was thinking to train the classifier first on many images containing only pure cultures of the respective classes (only yeast, only fungus A, only fungus b) and than check, if it is capable to segment the different components in microscope images of the mixed cultures. But I dont know how to do that.

Im very new to machine learning based image segmentation. I would be very glad if someone could help me with this.

Best, Ivan