ZEN Toolkits

AI Toolkit Life Science Image Analysis through Machine Learning

In life science applications, machine learning can exponentially increase the throughput of image analysis and reduce the risk of human error. This toolkit contains solutions for image denoising, image segmentation, and object classification.

Buy Now

Try Now

Improve every step of the image analysis workflow
Enable even new users to quickly gain proficiency
Import third-party machine learning models
30-day free trial

Enhance Every Step along the Analysis Workflow with Machine Learning

From optimizing raw images to identification and classification of objects

Image analysis is a multi-step process that requires raw image processing, identifying structures of interest by segmentation, and then classifying these structures based on their properties. ZEN AI Toolkit, powered by ZEN Intellesis, offers tools for each step of this complete workflow.

Intellesis Denoising optimizes the raw images.
Intellesis Segmentation identifies one or several classes of objects.
Intellesis Classification classifies these objects into meaningful subgroups.

Intellesis Classification then predicted the complete dataset

Classification of cellular phenotypes in U2OS + LLC2 cell culture. A few cells were manually assigned (left). Intellesis Classification then predicted the complete dataset (right).

Deep Learning Makes Automated Image Analysis Easier

A shift from manual programming enables even novice users

Conventional processing algorithms require the user to program and then fine tune the tools and parameters to achieve optimal analysis result. This requires an in-depth knowledge of image processing algorithms combined with programming knowhow and often lengthy trial-and-error experimentation to find the right tools and parameters for individual experiments. This complicated set of skills and time often put analysis algorithms out of reach for many researchers. With machine learning, the user teaches the algorithm the optimal result and the algorithm finds the optimal tools to solve the problem. This is a much faster process that is both intuitive and easy-to-use, even for novice users.

Labeling of an image to train a segmentation model (left) and automated image segmentation based on that model (right). Image courtesy of Yannick Schwab, EMBL Heidelberg

A Comprehensive, Unbiased View Delivers Better Performance

Looking beyond individual pixels in a digital image

The digital image acquired by a microscope is a set of pixel values. Conventional algorithms typically evaluate each pixel individually. A better understanding of the data provided from a digital image however comes from considering each pixel as well as their surrounding relationships with other pixels. This includes edges, gradients, textures, and shapes, as well as image artifacts such as background and shading. Machine learning delivers superior and more robust results by its ability to process large amounts of these data in an unbiased fashion without human bias or errors.

Technology Details for ZEISS Intellesis Segmentation

Unique methods to boost performance

Deep learning and machine learning combined: A VGG19 neural net feeding into the random forest classifiers

Combining the Strengths of Machine Learning and Deep Learning

Random Forest classifiers: The power of the majority vote
ZEISS Intellesis Segmentation works with a machine learning pixel classification algorithm known as random forest. This algorithm is based on decision trees that can classify pixels based on a high number of pixel features. Random forests use many of these trees and classifies pixels based on a majority vote. This leads to a very robust segmentation, requires minimal training data, and trains much faster than other algorithms.

A VGG19 neural net feeding into the Random Forest classifiers
Standard random forest classifiers may lack contextual awareness (the “neighborhood of a pixel”), relying only on standard image filters. By feeding the images into a VGG19 neural net and providing the feature maps to the random forest classifier, the speed of the random forest can be combined with the superior image recognition of deep neural nets. This leads to better contextual awareness and, consequently, better segmentation results.


Open architecture	ZEN workflow integration	Performance	Reliable algorithms
Import of deep-learning models from APEER and external tools (AI-ready) Exchange and sharing of models via open model format	Compatible with multi-dimensional datasets, including Z stacks and tiled images Seamless integration within Image Analysis Wizard and Bio Apps	Parallel and GPU computing Large data handling	Conditional Random Field algorithm (segmentation) Well-established open-source machine-learning algorithms, powered by Python, TensorFlow, ONNX, Scikit-Learn and Dask