Fig. 3.
Automatic expansion of reference sets. a, The automatic expansion scheme of reference sets to effectively select reference images that contribute most to the improvement of model performance. The unlabelled cell images from the training set D and an initial reference set Rn containing 1 reference image for each cell type K were provided in the first round. Images in D and Rn were cropped to 9 20 × 20 pixel sub-patches and processed by the trained feature encoder. The unlabelled cells were assigned with cell type K based on the Euclidean distance between embeddings of reference and unlabelled cells. The instance with the maximal Euclidean distance was selected for manual labelling and merged with Rn from the previous round to form the new reference set. In the experiment, the process was repeated 10 times. b, Examples of initial reference sets for each of the five datasets. c, Weighted F1-score on testing sets for the linear SVM classifier trained on the reference set generated by 3 different methods at each round of automatic expansion. The process was repeated 3 times with different initial reference images. The error bar indicates the standard error. The yellow horizontal line denotes the weighted F1-score achieved by the supervised classifier trained on 100% annotations.