Deep Learning
Artificial intelligence for image based applications

Artificial intelligence (AI) is the next frontier for imaging applications. DRVISION is at the forefront of AI-enabled technology for next-generation image analysis with Aivia, the first commercial image analysis software with a fully integrated end-to-end pipeline for deep learning.

 

Aivia provides a turnkey solution for applying pre-trained deep learning models for diverse imaging applications from restoration to segmentation and prediction. By providing some training images with ground truths (e.g. annotations), you can train a new model or update an existing one optimized for your application. Try Aivia today and experience the future of image analysis.

Deep learning 101

Aivia uses a wide range of AI techniques (e.g. random forest, support vector machines and deep learning) to solve image-related tasks. For the more demanding applications, Aivia uses a specific type of deep learning, known as fully convolutional neural networks (CNN).

CNNs are particularly well-suited for capturing non-linear relationships between large volumes of paired image sets (e.g. raw image and manually annotated image in the case of image segmentation tasks) thus allowing for a level of accuracy that rivals human experts (1,2). 

A neural network is composed of multiple artificial "neurons" organized in interconnected layers. Similarly to their biological counterparts, the neurons in artificial neural networks respond to specific stimuli, image patterns or features.

When stimulated, a neuron (both real and artificial) affects connected neurons in deeper layers until the output layer is reached where a prediction is made.

By comparing the prediction to the ground truth an error is calculated. The network is updated to minimize the error before a new round of learning (i.e. feeding a new input image to the CNN) takes place.

Initially, all neuron connections have equivalent weights and biases but during the process of learning the neuron paths that produce lower error values strengthen (i.e. connection weights are increased) while the ones with higher error values weaken.

The key processes responsible for learning in CNNs are gradient descent and back-propagation. Gradient descent helps determine how the weights and biases that govern the connections between different layers should be adjusted to minimize error.

Back-propagation implements the changes suggested by gradient descent, thus spreading the updates across the connected elements in the network.

The trained deep learning model includes the entire network with optimized weights and biases. It can be applied to "out of network" images (i.e. images that have not been used for training) similar to the ones used for training.

  1. Esteva et al., Nature, 2017

  2. Zeng, Wu and Ji,Bioinformatics, 2017

 

Model applications

In Aivia you can use deep learning models for three types of application: restoration, segmentation and prediction. You can apply a pre-trained model in Aivia or augment one of the models with your own data using Aivia Cloud.

Image segmentation

Segmentation of 3D electron microscopy (3DEM) datasets can be a tedious task for many researchers. Typically, scientists must go through hundreds of image slices and manually draw the boundary for each cell, which could take days to accomplish.

In Aivia, we have developed several pre-trained deep learning models based on different convolutional neural network architectures (DenseNet, UNet, 3D-UNet) to tackle EM image segmentation.

The deep learning model can applied to EM datasets like an image processing step. The model transforms the input image into a probability map of cell regions that can be easily segmented by threshold, enabling complete reconstruction of the image stack in minutes.

Image prediction

Image prediction takes a pair of images, such as phase contrast image and a corresponding fluorescent nuclei image, and creates a model for predicting the localization of the paired features. The model uses a UNet architecture to predict the localization of subcellular features (e.g. nuclei) from brightfield images. Aivia can create new image channel(s) for a desired feature that is virtually indistinguishable from fluorescence imaging of the feature.

The prediction model can be invaluable at prototyping imaging experiments - without using precious reagent. Additionally, the model works well in label-sparse or label-free environment, enabling you to amplify the data quantity to collect.

Apply one of the standard image analysis recipes in Aivia to the prediction output to segment and characterize the predicted objects. With prediction, you can extend your imaging experiment and analysis further than ever before.

 

Image restoration

Long-term imaging of live samples can be challenging for microscopists needing to balance limiting phototoxicity from light exposure without compromising image signal-to-noise ratio or spatial resolution. With deep learning image restoration, it is possible to obtain high-quality data while limiting the sample to low light exposure.

Using a Residual Channel Attention Network (RCAN), Aivia can enhance images acquired with extremely low light exposure to be indistinguishable from images acquired with optimal imaging conditions (i.e. much higher laser power and lower frame rate). Additionally, restoration with deep learning can retain (and recover) the spatial resolution of fine subcellular features (e.g. mitochondria, filaments). 

Using deep learning, you can perform long-term live cell imaging with significantly lower light exposure thus dramatically reducing photo-toxicity.

 

Accelerated training

Even with the most advanced GPU on the market, training a deep learning model can take many hours (or even days) per run. With Aivia Cloud, you can harness the power of cloud computing to accelerate the training process and obtain results in minutes.

Aivia Cloud gives you access to the state-of-the-art hardware offered by Google Cloud Platform for training deep learning models. Each training run can be parallelized across up to eight (8) NVIDIA V100 GPUs using Aivia Cloud. With Aivia Cloud, you can generate deep learning models quickly without the costly investment of purchasing and maintaining cutting-edge hardware.

Additionally, Aivia Cloud provides cloud-based storage for your data and remote access to powerful hardware running Aivia from anywhere - a complete analysis and visualization solution.

Processing
Storage
Remote access
Aivia Cloud
 

Resources

Aivia comes with several pre-trained deep learning models that integrates the most cutting edge neural network architectures available. Here are some of the solutions we have implemented in Aivia (with references):

 

Try Deep Learning in Aivia

For more information about pricing and try Aivia risk-free for 14 days, click on the buttons below.