Aivia Cloud (launched in parallel with Aivia 8) is a full-featured cloud-based image processing and visualization platform. It combines data storage, cloud computing, image processing and remote access. With Aivia Cloud, you can tap into virtually unlimited fast storage and state-of-the-art CPU and GPU hardware from anywhere with an internet connection.
Interface And Workflow
Aivia Cloud interface is shown below.
The Local Library box shows the content of the local computer folder linked to Aivia Cloud. You can adjust this setting by clicking on the Add Local Folder icon. The Cloud Library box shows the content of your folder in the Cloud. You can populate this folder by adding the chosen images from the Local Library to the Cloud Library. There are two ways to move the files between the Libraries - using Move Selected to Cloud/Local buttons and drag & drop.
The right hand side panel gives you the access to the cloud computed processes. To set up a new job, use the Create Job tab. To monitor the job progress, use the Progress Queue tab.
Training a Deep Learning model
The (minimum of two) pairs of images used to create your own Deep Learning model (Deep Learning Training) should be distributed between the Raw Input box and the Example box. To introduce a new pair of images to the new job, move your original image from the Cloud Library to the Raw Input box and do similar with your analyzed image - move it from the Cloud Library to the Example box. To move files you can right click on your image to chose the correct option, or drag and drop.
.png?version=2&modificationDate=1541189126464&cacheVersion=1&api=v2&width=377&height=150)
The pairs of images used for Training get annotated with a corresponding green number on the top left of the thumbnail (see the Aivia Cloud Interface picture). Please, make sure that your pairs are matching correctly before running the job.
The Training Progress Queue informs you about the progress of your cloud computed job. Uploading indicates the process of moving a file from the local computer to the cloud. Having your files in the Cloud Library reassures this part of the job is instant. Allocating indicates the process of distributing the job tasks to the virtually unlimited CPU and GPU hardware of the cloud. Training shows your Deep Learning model training progress. You can monitor the progress updates by clicking on the Status button. The Training Status provides you with real - time plots of the training accuracy as the job progresses.
.png?version=2&modificationDate=1541187043804&cacheVersion=1&api=v2&width=367&height=400)
The Training results are saved to the new folder in your Cloud Library. The folder name starts with "Training" and contains two copies of the same Deep Learning model file saved in two different formats: .hdf5e and .pbe.
.png?version=1&modificationDate=1541189419410&cacheVersion=1&api=v2&width=784&height=250)
Training files and Ground Truths
In Aivia Cloud creating your own deep learning model starts with two (sets of) images - raw data and example data (aka Ground Truth).
For image segmentation tasks the ground truth data is composed of your annotations which can be created using one of Aivia's image analysis recipes, using one of the intelligent semi-automatic tools in Aivia or can be imported from other software tools. The raw data are the images as they were acquired by the microscope.
For restoration tasks the ground truth data is a set of data which represents the absolute best quality images (for that set up/experiment) that can be generated. The raw data is composed of images created using imaging parameters which are very desirable (e.g. very low laser power or very fast acquisition) but that can only generate sub-par images.
For prediction tasks the ground truth data is often a set of images which show the localization of a fluorescently tagged protein or organelle. The raw data is often composed of images created using a different imaging modality (e.g. bright-field) which the user wishes to use more routinely for any number of reasons (e.g. cost, photo-toxicity, fluorescence labeling complexity).
To make sure that your Teaching will be successful, please make sure that your pairs of images:
- have the same bitness
- have the same XYZT dimensions (the current DL models work with a single time-point images only)
- have the same number of channels. Note that training on multi-channel data, will require a separate Ground Truth for each of the channels. You will also need to save each of your channels to a separate dataset, by right-clicking on it in Aivia and choosing Save
- have all the above mentioned characteristics the same for all the pairs within the Job.
Deep Learning model types
Aivia Cloud provides pre-trained Deep Learning models and supports custom creation of six types of Deep Learning models:
- 2D/3D image segmentation model
- 2D/3D image restoration model
- 2D/3D image prediction model
By checking the Transfer Learning box you can join the existing model of choice with the newly learned data.
Applying a Deep Learning model
The Applying mode allows you to apply your Deep Learning model to your images. Choose the correct Application mode and the model you want to use. Move the raw images you want to apply the model to, into the Raw Input box, and hit Create. The Progress Queue allows for the progress monitoring.
.png?version=1&modificationDate=1541191237857&cacheVersion=1&api=v2&width=283&height=400)
The Applying results are saved to the new folder in your Cloud Library. The folder name starts with the job time and date. To access your results, download them to your Local Library using the Move Selected to Local button.