Fig. 1
Unsupervised analysis of bone healing images. a The data analysis workflow starts with importing 37 images from a local folder. The images can be viewed in the Image Viewer widget (not shown) and are passed to the Image Embedder, which was set to use Google’s InceptionV3 deep network. We computed the distances between the embedded images and presented them as a dendrogram (b) with the Hierarchical Clustering widget. The clusters corresponded well to the time (days) post injury (D7 and D14), with a few exceptions. One such exception was a branch of two images highlighted in the dendrogram (b) and shown in the Image Viewer (2) (c). Image distances were also given to the multi-dimensional scaling widget (MDS), that also exhibits separation between bone healing samples at different times as depicted in different colors. Three representative MDS points from D7 and D14 were selected manually (data points with orange boundaries) (d) and the images are shown in the Image Viewer (1) (e). The two images highlighted in the dendrogram (b) were also passed to the MDS widget as a data subset. They are visualized as filled dots in this data projection (d) and they appear close to each other because of their similarity. This figure illustrates how a biologist may explore the data after clustering—first focusing on the misclassified samples and looking at the images and then selecting some of the best classified images as a point of reference for further exploration
