Segmentation of neural stem cells/neurospheres in unevenly illuminated brightfield images with shading reduction

Imaging neural stem cells/ neurospheres using low magnification brightfield modality results in uneven illumination effects across the field of view. Globally, the centre appears bright while the edges appear dark; locally, illumination varies across individual adjacent site images. Furthermore, neurospheres residing in the dark background regions have low signal:noise ratio. Altogether, they impose challenges for automated computer vision analysis. We propose a method which reduces the shading effects by formulating estimated global and local illumination models. Neurosphere detection is performed on the illumination corrected image using circular Hough transform followed by the removal of false detections. Experimental results validated on 480 images, containing 738 neurospheres, show that our method achieves a detection score of 90.76% precision and 81.17% recall. Our method allows for the automated analysis of neurospheres in high content screens, and could be used to study the effects of drugs on neurospheres, using brightfield microscopy and without fluorescence staining.