Cellular Segmentation of Bright-field Absorbance Images Using Residual U-Net

Cellular segmentation is one of the essential processes for cytometric analysis. Bright-field microscopy (absorbance) is a widespread technique among all optical microscopy approaches. However, due to the low contrast in absorbance images, membranes of cells are not easily discernible. Convolutional Neural Network (CNN) based deep learning approaches have successfully been applied in many medical image processing applications. In this paper, we propose a residual network-based U-Net for segmentation of cell region in cell absorbance images of human-induced pluripotent Retinal Pigment Epithelial stem cells (iRPE). Additionally, we use a weighted sum of multiple loss functions, namely, Binary Cross-entropy loss, Dice loss, and inverted Dice loss, to optimize the network weights. With this proposed approach, we achieve an average Dice Coefficient of 0.8366 on the validation data split for the intended task.