To make our methods more readily available and duplicable, we have coded a set of Colab notebooks on which users can both test and train the neural networks used in this study ( Colab notebooks are designed to run using a hosted runtime on the Google cloud service. This allows users access to all of the necessary coding libraries, as well as GPU compute power, without having to install and setup the environment locally. Notebooks were created for Mask R-CNN and ANCIS, as well as the UNet used herein for noise detection. A separate notebook for StarDist was not created, since a good Colab notebook focusing on nucleus segmentation already exists for this network, and can be used with our pretrained weights [40]. Our modifications for post processing predictions and resizing images were also included in the code, as optional parameters. We have also included links to our pre-trained weights for each network using our super-resolution image sets.

The StarDist distribution applied here was version 0.6.0, downloaded from the author’s GitHub page [20]. Parameters used included a batch size of 2 and a training patch size of 128 × 128. The number of rays for the star distributed polygons was set at 32, the author recommended value. During prediction, the non-maximum suppression (nms) threshold was set to 0.3 and the probability threshold to 0.5. Additional settings were left to their default values, including binary cross-entropy training loss function and mean average error polygon distance loss function parameters.

Our tested distribution of the ANCIS code was also downloaded from the author’s GitHub page [14]. Network parameters used for all training and testing included a nms threshold of 0.7, confidence threshold of 0.5 and segmentation threshold of 0.5. Thresholds were selected based on previous experiences with image segmentation networks, as well as findings reported in the literature. The batch size was limited to 2 by the GPU capacity. The maximum detectable instance limit was set greater than the number of instances in any of the training or test images, we used 400, since we did not wish to limit segmentation in this way. Anchors were left at defaults values, along with default cross entropy loss function.

Matterport’s broadly disseminated Mask R-CNN distribution, version 2.1, was implemented in these experiments [41]. Threshold parameters and maximum instance limits were matched to those reported above for ANCIS. In addition, we applied some of the parameters set by Waleed Abdulla in his nucleus segmentation example code [41]. Anchors per image were set to 64, and anchor scales were set at (8, 16, 32, 64, 128). We did not implement mini masks, however, and our batch size was limited to 1. The network backbone used was resnet50, and binary cross entropy loss function as used. Training was initialized on the pre-trained coco dataset weights, downloaded from the code author’s website [41].

Note: The content above has been extracted from a research article, so it may not display correctly.

Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.

We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.