Fig. 4. Model training started with a relatively small dataset and its performance increased by annotating more objects.

a A schematic of active learning process. b After using the active learning approach, per day mean average precision (mAP) has been improved drastically. Here in each iteration, 250 new tiles of mostly rare cellular and non-cellular objects are selected to be annotated and then merged to the current dataset. Without using active learning, all objects in the new tiles are annotated by the hematopathologists, but by using that, the model trained from the current dataset is run on the new tiles to detect and classify the objects and then hematopathologists review the results and confirm or modify them. In each iteration, the new annotated tiles are merged with the current dataset. The model will be trained on this new dataset and the next iteration will be started.