Cascaded 3-Stage Nuclei Segmentation Using U-net, Faster-RCNN and SegNet for Higher Precision

Though there had been much new deep learning based nuclei segmentation architectures available for various applications made by researchers in the last decade, most of them are specific for particular nuclei types of imaging modalities. In this work, we are proposing a step-wise pipeline for nuclei segmentation comprising Unet based difference image extraction, Faster-RCNN for nuclei detection on difference image, and SegNet for Nuclei segmentation from the detected nuclei inside bounding boxes. Our proposed algorithm provides significant improvement when it comes to Mean Average Precision vs Mask-RCNN and provides a general framework where advantages of transfer learning could be deployed easily. Due to the use of cascaded approach, each stage has a single cost function to minimize which helps to reach the global optimum with a limited number of training images and could be separately modified as necessary.