We compare several deep neural network-based image processing techniques for the analysis of cellular behavior of cells cultured directly on Lab-on-CMOS devices. Lab-on-CMOS devices are typically opaque and use integrated circuits to implement functionality, so they must be observed using reflection mode microscopy and have prominent background features, significantly increasing the difficulty of the cell segmentation task. Relative to previous approaches based on morphological filtering, the neural-net based approaches improve the intersection-over-union metric for image segmentation from 57% to 85%.