Figure 4
Unsupervised learning identifies different blobs. (A) The unsupervised learning framework to build the blob identification model based on datasets. Training phase: we used the cells from both populations of the first three experiments (Fig. 1B) to build the learning model using the unsupervised clustering. The cells are divided into ROIs. A multi-threshold network analysis for each ROI is employed to filter-out the noisy blinks and find the clustered nodes. The blobs are generated from the clustered nodes using the mean shift algorithm. A new set of features are extracted from each blob and fed into the unsupervised clustering (X-means) to learn the different groups. The groups from PC3 and PC3-PTRF populations are matched using the similarity analysis to identify the groups’ types. The matched groups are used to label the blobs on the cells. Testing phase: we used the built model to identify the blobs of the cells from experiment 4. The cell is passed via the space division to get the ROIs. The multi-threshold analysis is applied to filter-out the noisy blinks and return the clustered nodes. The blobs are generated using the mean shift algorithm. The same set of features is extracted for each blob. Each blob feature vector is tested against the centroid feature vector of the learned groups. The closest distance is the most similar group to this blob. The blobs are finally labeled based on the similarity of their feature vector with groups’ centroids. (B) After filtering, blob-level feature analysis and segmentation identifies 2 groups (P1, P2) in PC3 and 4 groups (PP1, PP2, PP3, PP4) in PC3-PTRF by unsupervised learning. One-to-one group matching (box) with distances among feature vector of groups centers used as the similarity measure (closer groups are more similar). (C) Each group of blobs (S1 or S2 scaffolds or caveolae) is extracted and shown as different channels. Graph shows percent distribution of blobs in PC3-PTRF and PC3 cells.
