In the post-processing, the raw predictions are binarized using a global threshold of 0.5, and the centroids of predicted beads are calculated. However, depending on the resolution and the seeds’ distances the number of missing detections may also increase. This enlarges the seeds and reduces the number of missing seed detections. To get a more robust seed detection, a morphological dilation with a cross-like mask can be applied to the upsampled training label images.
To improve the generalization ability of the trained model, specific training data augmentations are applied online during training. Thereby, the traditional upsampling utilizes prior knowledge enforcing the network to work at reasonable scales. Thus, for the bead detection, an adapted U-Net trained with a combination of binary cross-entropy and Dice loss is used ( Milletari et al., 2016 Ronneberger et al., 2015). Due to the upsampling preprocessing, centers of touching beads do not share an edge anymore, and the bead detection can be treated as a semantic segmentation task. Furthermore, it eases the annotation of a low-resolution dataset that is required for supervised learning ( Fig. 1d). The bilinear upsampling, hereby, enables distinguishing touching beads ( Fig. 1b and c). BeadNet also provides the automatic calculation of evaluation measures and overlays of the detection results.īeadNet consists of an upsampling preprocessing step for low-resolution images and a deep learning-based bead detection with subsequent counting.
The upsampling step and the training data generation are often missing in toolboxes, leaving users alone with the upsampling, the normalization and the creation of new training samples. In addition, BeadNet provides a graphical user interface for easy use, e.g.
BEAD TOOL 4 ADD TEXT SOFTWARE
The software combines a crucial upsampling preprocessing with a deep learning-based detection. To our knowledge, this is the first annotated 2D dataset and software tackling the problem of counting beads in low-resolution images with high accuracy. Despite its simple use, BeadNet enables a nearly error-free bead detection on a manually annotated low-resolution bead dataset. For domain adaptation, biologists only need to create new training images using a built-in functionality and annotate them using our labeling tool. Thus, we designed the bead detection and counting software BeadNet to avoid parameter adjustments. Even for experts, it is often easier to annotate new images instead of adjusting parameters for low-resolution and low signal-to-noise-ratio data. In addition, the parameters of sophisticated analysis pipelines need to be adjusted in case of changes in the experimental design.
BEAD TOOL 4 ADD TEXT MANUAL
The complexity of bead detection in low-resolution images may lead biologists to either use an inefficient and user-biased manual quantification or to use a user-friendly but inaccurate method. The predictions of BeadNet are shown in ( e) (white in gray ground truth) The white 2 × 2 px seeds in ( d) are enlarged (gray) for the calculation of the evaluation metrics. The upsampled test image is shown in ( c). The 32 × 32 px test image in ( b) is taken from the red fluorescent channel of (a). ( a) A maximum intensity projection of red fluorescent beads (additional green fluorophore for beads outside of cells) and blue fluorescent cell nuclei.
Exemplary application case and detection results of BeadNet.
0 kommentar(er)
