Supervised learning for cell classification is one of the most used approaches on different studies. However, due to lack of labelling datasets provided by experts, and the small number of images per dataset, the usage of unsupervised learning would be a better approach. This work reports a study done on a two-stage unsupervised classification of cell health: i) phase one: divide the dataset into two main groups healthy and unhealthy, and ii) phase two: divide healthy group into two smaller clusters and unhealthy group into three different clusters. The groups are defined following the description of the cell health ISO STANDARD for invitro cytotoxicity evaluation. Unsupervised learning is done based on image features and the labelling of the clusters is done after it. K-means algorithm was used for clustering and two different datasets were tested. The second dataset had the best performance and the correct labelling of the clusters.