I'll approach the following problem next week and would love to get some advice. I am considering modeling it as a regression problem for x,y,r or as an object detection problem.

I have microscopy images coming from multiple sources(different contrast, brightness, zoom, focus), and each image has a circle in it that I want to detect.

An image can have multiple circles in it(of which I am interested in a specific one) and it can also have a single circle. Ultimately, I am interested in the area of that circle.

I have around 4k of annotated images(bounding box around the circle) from 3 sources, I have gotten good results with using hugh transform but it doesn't generalize properly to the other image sources since they have different amounts of circles in them.

I am starting to think that using deep learning with heavy augmentations(contrast, brightness, etc...) will help me generalize to all the different sources.

Any input would be appreciated.