Algorithms ease data-heavy image analysis
Slice of test mouse cell with corresponding manual segmentation and output of new neural network with 100 layers [Ekman, Larabell, National Center for X-ray Tomography]
Mathematicians at US-based Berkeley Lab have developed a new approach to machine learning aimed at experimental imaging data.
Rather than relying on the tens or hundreds of thousands of images used by typical machine learning methods, a new neural network architecture 'learns' more quickly and requires far fewer images.
Pioneered by Daniël Pelt and James Sethian of Berkeley Lab’s Center for Advanced Mathematics for Energy Research Applications, so-called 'Mixed-Scale Dense' convolution neural network, requires fewer parameters than traditional methods, converges quickly, and can 'learn' from a remarkably small training set.
The new algorithm is being used to extract biological structures from cell images, and is set to analyse data across a wide range of research areas, critical as experimental facilities generate higher resolution images at ever-higher speeds.
“In many scientific applications, tremendous manual labour is required to annotate and tag images, but it can take weeks to produce a handful of carefully delineated images,” points out Sethian. “Our goal was to develop a technique that learns from a very small data set.”
“The breakthrough resulted from realizing that the usual downscaling and upscaling that capture features at various image scales could be replaced by mathematical convolutions handling multiple scales within a single layer,” adds Pelt.
Many applications of machine learning to imaging problems use deep convolutional neural networks in which the input image and intermediate images are convolved in a large number of successive layers, allowing the network to learn highly nonlinear features.
However, the new “Mixed-Scale Dense” network architecture calculates dilated convolutions as a substitute to scaling operations to capture features at various spatial ranges, employing multiple scales within a single layer, and densely connecting all intermediate images.
Crucially, the new algorithm achieves accurate results with few intermediate images and parameters, eliminating both the need to tune hyperparameters and additional layers or connections to enable training.
Tomographic images of a fibre-reinforced mini-composite, reconstructed using 1024 projections (a) and 128 projections (b). In (c), the output of an Mixed-Scale Dense convolution neural network with image (b) as input is shown. A small region indicated by a red square is shown enlarged in the bottom-right corner of each image.
Pelt and Sethian are now applying the algorithm to a host of new areas, such as fast real-time analysis of images coming out of synchrotron light sources and reconstruction problems in biological reconstruction such as for cells and brain mapping.
“These new approaches are really exciting, since they will enable the application of machine learning to a much greater variety of imaging problems than currently possible,” says Pelt. “By reducing the amount of required training images and increasing the size of images that can be processed, the new architecture can be used to answer important questions in many research fields.”
The algorithm can be accessed at this web portal: “Segmenting Labeled Image Data Engine (SlideCAM)” as part of the CAMERA suite of tools for DOE experimental facilities.
Research is published in PNAS.