Hi! I'm wondering how model training works? I am currently trying to train a model in cellpose 3.0.8 to identify all muscle fiber cells in a cross section with IF Laminin staining (stains extracellular matrix-images attached below). I first tried to train the model based on one cropped-out muscle group. However, after about 4 different cropped images, the model was spitting out an ROI coverage not nearly as great as just using cyto3- especially when applying the models to an entire cross section with all muscle groups. Is this based off of the set calibration? Does the size of the image (in terms of zoom in/out and #cells per image) matter when it comes to model accuracy?
I also noticed that when you train the model, the initial model used is the cyto3 for each. Are each of my models that I have made and corrected based solely on cyto3 and not the previous model I am trying to build? If the models are being built on top of one another (as I hope), how does this work and is there a way to see what models the new model will be based off of? Otherwise, how do I make a model that is much more accurate in identifying muscle fiber cells? Is there a repetition range I should shoot for during training?
I am also noticing that the draw ROI bug is still present with cellpose 3.0.8. The red circle disappears but the mouse tracker trail stays and does not leave unless I exit the program/ add a new image. When this happens, should I save my model with the image I am correcting right then or start over with corrections? (does the program identify the trail as draw ROI or can it not recognize it?)
Hi! I'm wondering how model training works? I am currently trying to train a model in cellpose 3.0.8 to identify all muscle fiber cells in a cross section with IF Laminin staining (stains extracellular matrix-images attached below). I first tried to train the model based on one cropped-out muscle group. However, after about 4 different cropped images, the model was spitting out an ROI coverage not nearly as great as just using cyto3- especially when applying the models to an entire cross section with all muscle groups. Is this based off of the set calibration? Does the size of the image (in terms of zoom in/out and #cells per image) matter when it comes to model accuracy?
I also noticed that when you train the model, the initial model used is the cyto3 for each. Are each of my models that I have made and corrected based solely on cyto3 and not the previous model I am trying to build? If the models are being built on top of one another (as I hope), how does this work and is there a way to see what models the new model will be based off of? Otherwise, how do I make a model that is much more accurate in identifying muscle fiber cells? Is there a repetition range I should shoot for during training?
I am also noticing that the draw ROI bug is still present with cellpose 3.0.8. The red circle disappears but the mouse tracker trail stays and does not leave unless I exit the program/ add a new image. When this happens, should I save my model with the image I am correcting right then or start over with corrections? (does the program identify the trail as draw ROI or can it not recognize it?)
Muscle Group Cropped.pdf