Extended Data Fig. 3: Detection performance and ablation studies on contrastive learning and positive unlabeled components.
From: MiLoPYP: self-supervised molecular pattern mining and particle localization in situ
Detection accuracy metrics were evaluated for in vitro and in situ datasets EMPIAR-10304 (a) and EMPIAR-10499 (b), respectively. Precision, recall and F1 scores were compared against crYOLO-3D and template matching (top row). We also measured detection performance using F1-scores when using increasing numbers of annotations during training (middle row). On the bottom row, we present mean, minimum and maximum average F1 scores for the ‘complete model’ including the contrastive (CR) and positive unlabeled (PU) learning components (blue), without the CR component (orange), and without the PU component (green), demonstrating the efficacy of the model. The upper bars represent the maximum minus the mean, and the lower bars represent the mean minus the minimum (sample size is 10 for EMPIAR-10304 and 5 for EMPIAR-10499).
