Main content

Home

Menu

Loading wiki pages...

View
Wiki Version:
![enter image description here][1] <br> Candescence version 1.0 [software]<br> [Varasana][2] version 1.0 [training data]<br> Candescence.io version 1.0 [web interface]<br> Bettauer et al. (2021) [bioRxiv paper](https://www.biorxiv.org/content/10.1101/2021.06.10.445299v1) This work is under a Creative Commons Zero v1.0 Universal license. ---------- This sites serves as a repository for the **Candescence** software project. Candescence uses deep learning approaches to tackle different challenges related to the considerable array of morphologies for the opportunistic human fungal pathogen *Candida albicans*. This is a [short presentation](https://youtu.be/A6VpKFKpZA4) (12 minutes) of Candescence from our talk at Candida and Candidiasis 2021. Some movies of the GAN can be found [here](https://osf.io/qdxbp/wiki/Movies%20of%20GAN-based%20trajectories/) Candescence can identify C. albicans cells in microscopy images (DIC to date although we are working on other image types). This is constructed using a FCOS, a deep learning technique that efficiently locates objects and classifies each object in an image. Candescence also builds models of different morphologies and transitions between morphologies using a Generative Adversarial Network (GAN). Candescence is trained using the **Varasana** dataset. We are working on web tool to more easily use Candescence. It will be located here [candescence.io][3]. We are currently struggling though to make this work well given all the dependencies with other software packages including MMDETECTION. Until this is available we provide a well-annotated Jupyter notebook to run the FCOS classifier and to explore the GAN models (provided in the Github repository). We are also happy to assist you to run your images. This repository links to the following: 1. A Github respository which houses both the software used to train Candescence and the software that supported the manuscript. This is R and Python code. 2. Google Drive-based documents that include the manucript, presentations, figures, movies and a (Zotero-based) bibliography. 3. All of the images (train, validation, test) for Varasana. Initially, these images and items 4 and 5 described below are available at the [Varasana temporary page][4]. 4. The annotations of the microscopy data for the training and validation components of the training set. The are also available at the Varasana temporary page. 5. If we receive approval from the [EBI BioImage archive][5] to house all of the microscopy data, we will move the images to this location. Currently both users and developers are required to install [MMDETECTION][6] on their machine. We build on top of this Python package. There are additional packages used by MMDETECTION that you will need to build the appropriate conda environment. The R code here uses version 4.0.4. There are installation instructions at the github site. The **Varasana temporary** data repostiory contains 5 subdirectories. 1. **GAN** provides some files related to the training in different contexts. 2. **varasana** has subdirectories: (i) **all-original-images**. These are the original DIC images that have not been re-scaled (800x800) for training. The raw subdirectory are tif files directly from the microscope. dic contains bmp versions. The names of these files can be found in Supplemental Table 1. (ii) **original-labelbox-annotations** contains a json file that is produced by the Labelbox software. If you would like to develop your own dataset or modify our labels, we recommend the [Labelbox](https://labelbox.com/) tool. (iii) **test** contains the bmp images (rescaled to 800x800) restricted to the test set (as per Supp Table 1). (iv) **train-validation** contains a series of pkl files for each grade in both the train and validation datasets. Note that although the pkl files in train point to a set of images disjoint from validation. However the subdirectories train and val contain all images (they are not disjoint). 3. The **performance** directory contains the results of our experiments with the Candescence FCOS model. There are two subdirectorie. The training-of-model directory contains the output and reports generated buy the MMDETECTION-based FCOS. The analysis-of-model directory contains the output generated by our python and R based code. This is an extended version of the results presented in the paper. 4. The **production** directory contains the pth files describing our deep learning models. In particular, the candescence_version_1.0 subfolder contain both the pth file of the model parameters and the MMDETECTION configuration file (.py). We will add new models to this folder as they become availalbe. We would very much appreciate your comments, criticisms, suggestions, ideas for extensions, and random ad hominem attacks. You can leave them at the [To Do](https://osf.io/qdxbp/wiki/To%20do%20list/) list or contact us direclty at [hallett.mike.t@gmail.com][7] P.S. The Candescence logo (above) is an 80s throwback thing, combining the aesthetic of that classic 1984 album Meat is Murder by The Smiths, with a hint of Joy Division's Love Will Tear Us Apart. ![enter image description here][8] Project led by the [Hallett lab][9]. ---------- [1]: https://files.osf.io/v1/resources/qdxbp/providers/osfstorage/60abd0eb0b6c6902339b782c?mode=render [2]: http://csfg-algonquin.concordia.ca/~hallett/candescence [3]: http://candescence.io [4]: http://csfg-algonquin.concordia.ca/~hallett/candescence [5]: https://www.ebi.ac.uk/bioimage-archive/ [6]: https://mmdetection.readthedocs.io/en/latest/#ne.tp://candescence.io [7]: http://mailto:hallett.mike.t@gmail.com [8]: https://files.osf.io/v1/resources/qdxbp/providers/osfstorage/60abd5f10b6c69022a9c0b4e?mode=render [9]: https://mikehallett.science