Region of Interest Localisation of Hematopoietic Stem/Progenitor Cell Images

Abstract

Image classification using deep learning has been widely implemented, primarily in medical imaging. However, features and focus regions are extracted by the network, becomes a black box mystery in the feature extraction layer during network training, unlike conventional feature extraction approaches, where various methods can extract image features. Regardless, traditional image feature extraction is laborious to find the most suitable algorithm. It takes time to meet the significant image features before classification and final image localisation, especially for the microscopic images. Therefore, a method to localise the region of interest (ROI) in vitro of the colony-formation unit (CFU) of hematopoietic stem/progenitor cell (HSPC) using gradCAM through deep learning, approaches have been proposed. This work comprises three main phases: CFU data preparation, convolutional neural network (CNN) pre-trained networks and localisation of the ROI. The proposed method has successfully localised the ROI of the CFU HSPC using gradCAM through a deep neural network with 87.5% sensitivity performed by DarkNet19. The finding of this work can be used as a baseline for future CFU HSPC classification that focuses on the CFU region.