Fig. 4: LACSS model accuracy trained with LIVECell dataset.
From: Efficient end-to-end learning for cell segmentation with machine generated weak annotations
Models were trained either under full-supervision using the original single-cell annotation (supervised, solid line), under semi-supervision using synthetic annotation but without auxnet (dotted line), or under semi-supervision with boundary-defining auxnet (dash line). AP values were computed for eight cell lines at increasing IOU thresholds from 50% to 95%. The results showed significant improvements in model accuracy by introducing auxnet in the training for models trained with semi-supervision.
