Figure 3.
YOLOv2 object detection performs well in a number of domains. (A–D) YOLOv2 detection accuracy is consistent across images of lower digital resolution but is sensitive to optical resolution. (A) COS-7 cells stained for nuclear pore and imaged using a 1.4 NA 100× objective. (B) Same image as in A resampled at 50% of pixel resolution. (C) Same image as in A but resampled at 20% of pixel resolution. Scale (A–C) is 20 µm. (D) The accuracy of prediction on the 0.5 (B) and 0.2 scaled dataset (C) stays relatively high compared with normal resolution (1.0, 100×) when compared across three independent datasets (n = 3, mean ± SD). (E) COS-7 cells stained for nuclear pore and imaged with 40× 0.55 NA objective (scale bar, 20 µm). The dotted frame represents image area of equivalent physical scale of images A–C i.e., 133.12 × 133.12 µm. (F) COS-7 cells stained for nuclear pore and imaged with a 10× 0.45 NA objective (scale bar, 100 µm; inset, 20 µm). The white dotted inset is a zoom region on the dotted frame and represents an equivalent physical dimension of A (i.e., 133.12 × 133.12 µm); the red dotted frame represents same physical dimension as E (i.e., 200 × 200 µm). (G) Graph showing optical resolution is critical for performance of YOLOv2 when used to evaluate images collected on different microscopes with different objective types (40× and 10×; n = 3, mean ± SD). (H–J) Multiclass cellular detection in HEK peroxisome and erythroid DAPI datasets. YOLOv2 can be used to discretely identify cells with specific visual phenotypes within a single image. (H) HEK cells with varying levels of GFP-SCP2 expression, with either punctate fluorescence or a low level of diffuse nonpunctate fluorescence. (I) Erythroid nuclei stained with DAPI, highlighting either single/multinucleate cells that are healthy or in a state of apoptosis characterized by a blebbed appearance. In both images (H and I), white dashed ROIs represent ground-truth annotations used for training. ROIs with a star in close proximity represent the subset of annotations that were labeled positive for a phenotype of interest in the training. Blue ROIs represent predicted regions from model trained to recognize all cells in the image, whereas the green ROIs represent prediction for model trained to recognize a specific subset of cells present. (H) Output of an experiment with ROI predictions from a classifier trained to recognize all the cells in the image, while the second model (blue ROI) was trained to only recognize the cells exhibiting punctate fluorescence. (I) Output of an experiment where one network was trained to recognize only heathy single/multinucleate cells (green ROI), whereas the second model was trained to recognize all cells, including apoptotic cells (blue ROI). In perfectly classified images, green regions should only appear next to annotation regions with a green star (I has some wrongly classified regions). Scale bars represent 20 µm in both images. (J) Graph summarizes the AP measured with respect to the different conditions in H and I. Detector is capable of recognizing phenotypic subsets; however, the accuracy drops (in these experiments) when the network is trained to recognize a subset of cells (white bars) rather than cells in each image (filled bars; n = 3, mean ± SD).

YOLOv2 object detection performs well in a number of domains. (A–D) YOLOv2 detection accuracy is consistent across images of lower digital resolution but is sensitive to optical resolution. (A) COS-7 cells stained for nuclear pore and imaged using a 1.4 NA 100× objective. (B) Same image as in A resampled at 50% of pixel resolution. (C) Same image as in A but resampled at 20% of pixel resolution. Scale (A–C) is 20 µm. (D) The accuracy of prediction on the 0.5 (B) and 0.2 scaled dataset (C) stays relatively high compared with normal resolution (1.0, 100×) when compared across three independent datasets (n = 3, mean ± SD). (E) COS-7 cells stained for nuclear pore and imaged with 40× 0.55 NA objective (scale bar, 20 µm). The dotted frame represents image area of equivalent physical scale of images A–C i.e., 133.12 × 133.12 µm. (F) COS-7 cells stained for nuclear pore and imaged with a 10× 0.45 NA objective (scale bar, 100 µm; inset, 20 µm). The white dotted inset is a zoom region on the dotted frame and represents an equivalent physical dimension of A (i.e., 133.12 × 133.12 µm); the red dotted frame represents same physical dimension as E (i.e., 200 × 200 µm). (G) Graph showing optical resolution is critical for performance of YOLOv2 when used to evaluate images collected on different microscopes with different objective types (40× and 10×; n = 3, mean ± SD). (H–J) Multiclass cellular detection in HEK peroxisome and erythroid DAPI datasets. YOLOv2 can be used to discretely identify cells with specific visual phenotypes within a single image. (H) HEK cells with varying levels of GFP-SCP2 expression, with either punctate fluorescence or a low level of diffuse nonpunctate fluorescence. (I) Erythroid nuclei stained with DAPI, highlighting either single/multinucleate cells that are healthy or in a state of apoptosis characterized by a blebbed appearance. In both images (H and I), white dashed ROIs represent ground-truth annotations used for training. ROIs with a star in close proximity represent the subset of annotations that were labeled positive for a phenotype of interest in the training. Blue ROIs represent predicted regions from model trained to recognize all cells in the image, whereas the green ROIs represent prediction for model trained to recognize a specific subset of cells present. (H) Output of an experiment with ROI predictions from a classifier trained to recognize all the cells in the image, while the second model (blue ROI) was trained to only recognize the cells exhibiting punctate fluorescence. (I) Output of an experiment where one network was trained to recognize only heathy single/multinucleate cells (green ROI), whereas the second model was trained to recognize all cells, including apoptotic cells (blue ROI). In perfectly classified images, green regions should only appear next to annotation regions with a green star (I has some wrongly classified regions). Scale bars represent 20 µm in both images. (J) Graph summarizes the AP measured with respect to the different conditions in H and I. Detector is capable of recognizing phenotypic subsets; however, the accuracy drops (in these experiments) when the network is trained to recognize a subset of cells (white bars) rather than cells in each image (filled bars; n = 3, mean ± SD).

This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal