Implementation of transfer learning for the segmentation of human mesenchymal stem cells—A validation study

Abstract

IntroductionStem cell therapy has been gaining interest in the regeneration rather than repair of lost human tissues. However, the manual analysis of stem cells prior to implantation is a cumbersome task that can be automated to improve the efficiency and accuracy of this process. ObjectiveTo develop a Deep Learning (DL) algorithm for segmentation of human mesenchymal stem cells (MSCs) on micrographic images and to validate its performance relative to the ground truth laid down via annotation. MethodologyPre-trained DeepLab algorithms were trained on annotated images of human MSCs obtained from the open-source EVICAN dataset. This dataset comprises of partially annotated images; a limitation that is overcome by blurring backgrounds of these images which consequently blurs the unannotated cells. Two algorithms were trained on the two different kinds of images from this dataset; with blurred and normal backgrounds, respectively. Algorithm 1 was trained on 139 images with blurred backgrounds and algorithm 2 was trained on 37 images from the same dataset with normal backgrounds to replicate real-life scenarios. ResultsThe performance metrics of algorithm 1 included accuracy of 99.22%, dice co-efficient of 99.66% and Intersection over Union (IoU) score of 0.84. Algorithm 2 was 96.34% accurate with dice co-efficient and IoU scores of 98.39% and 0.48, respectively. ConclusionBoth algorithms showed adequate performance in the segmentation of human MSCs with performance metrics close to the ground truth. However, algorithm 2 has better clinical applicability, even with smaller dataset and relatively lower performance metrics.