Nuclei classification gained immense interest among researchers to build computer-assisted diagnosis (CAD) tools. With the advancement of image processing and pattern recognition techniques, CAD has been widely used to help medical professionals to interpret medical images. Digital pathology, as a vital aspect of CAD application, is gaining a lot of attention from both image analysis researchers and pathologists because of the advent of whole-slide imaging. The potential of digital pathology span over a wide range of applications such as segmentation of desired regions or objects, counting cancer cells, recognizing tissue structures, classifying cancer grades, a prognosis of cancers, etc. In this research study, we propose a deep learning-based approach in order to detect and segment different nuclei present in diverse cell images, followed by postprocessing method to identify the overlapping nuclei in the images. We have added pre-processing steps for processing the raw image which includes rescaling, resizing, converting to high intensity background, data augmentation and set of morphological operations. Pre-processed images are then used for pixel level classification and segmentation. We have implemented U-Net architecture, a Convolutional Neural Network (CNN) based neural network for semantic segmentation and added few steps that processes segmented image to identify overlapping nuclei and mark all nuclei. Our study shows that with the help of data pre-processing step, we were able to increase the efficiency of the model and post classification steps helps to identifying overlapping nuclei.
The proposed methods have achieved around 0.96 of average dice coefficient as the accuracy of pixel level classification