NeuroTrALE: Leveraging Machine Learning for Advanced Brain Mapping and Alzheimer’s Research

In late 2023, the U.S. Food and Drug Administration (FDA) approved the first drug aimed at slowing the progression of Alzheimer’s disease. While this marks a significant step forward in the battle against a condition that affects millions worldwide, understanding and combating Alzheimer’s and other neurological disorders remains a complex challenge. Machine learning is now emerging as a key tool in this effort, offering new ways to analyze and interpret vast amounts of brain imaging data.

Lars Gjesteby, a technical staff member and algorithm developer at MIT Lincoln Laboratory’s Human Health and Performance Systems Group, describes the challenge: “Understanding the human brain at a cellular level is one of neuroscience’s toughest problems. Machine learning can significantly aid this process by providing tools to handle large-scale brain data more efficiently.” High-resolution, networked brain atlases could improve our understanding of neurological disorders by revealing structural and functional differences between healthy and diseased brains. However, creating these atlases requires advanced tools that can process and visualize massive imaging datasets, an area where machine learning has great potential.

A networked brain atlas is essentially a detailed map of the brain that links its structure to its functions. Building these atlases involves processing and annotating large amounts of brain imaging data, where each axon — the fibers connecting neurons — must be meticulously traced, measured, and labeled. Traditional methods, which rely on desktop software or manual techniques, are not equipped to handle human brain-scale data efficiently, resulting in researchers spending countless hours manually processing raw data.

Gjesteby is leading the development of the Neuron Tracing and Active Learning Environment (NeuroTrALE), a cutting-edge software pipeline that combines machine learning, supercomputing, and user-friendly interfaces to streamline brain mapping. NeuroTrALE automates much of the data processing and provides an interactive platform where researchers can manipulate data to identify, filter, and search for specific patterns. This innovation significantly enhances the efficiency of creating brain atlases.

One of the standout features of NeuroTrALE is its use of a machine learning technique known as active learning. Unlike traditional machine learning, where algorithms passively learn from a set dataset, active learning allows the algorithm to interactively query users to correct errors in real-time. This dynamic approach ensures that the algorithm improves its performance with each new dataset it encounters, allowing for more accurate and efficient data labeling.

Michael Snyder, from the Homeland Decision Support Systems Group at MIT Lincoln Laboratory, illustrates the concept: “Think of it like examining an X-ray of a tangled ball of yarn with all its overlapping lines. It’s hard to tell if a strand is bending or if two different strands are crossing over each other. With NeuroTrALE’s active learning, users can correct the algorithm a few times, teaching it to follow the correct path moving forward. Without NeuroTrALE, researchers would have to manually trace each axon every time.” This capability allows NeuroTrALE to dramatically reduce the time required to process large datasets. In fact, using NeuroTrALE, the team achieved a 90% reduction in processing time for 32 gigabytes of data compared to conventional AI methods.

Moreover, the NeuroTrALE team has shown that substantial increases in data volume do not lead to proportionate increases in processing time. For instance, a 10,000% increase in dataset size resulted in only a minimal increase in processing time, thanks to the parallel computing capabilities that distribute tasks across multiple GPUs.

Benjamin Roop, an algorithm developer on the project, underscores the impact of this technology: “With around 86 billion neurons and 100 trillion connections in the human brain, manually labeling all axons would be a monumental task. NeuroTrALE’s automation capabilities mean we can generate detailed brain maps, or connectomes, for multiple individuals, opening new avenues for large-scale studies of brain diseases.”

The NeuroTrALE project began as a collaboration between Lincoln Laboratory and MIT’s Chung Lab, supported by internal funding. By 2022, the Chung Lab was using NeuroTrALE to produce meaningful insights, including findings published in Science. They used NeuroTrALE to analyze cell density in the prefrontal cortex of Alzheimer’s patients, revealing lower cell density in affected areas and identifying where harmful neurofibers tend to tangle in diseased brain tissue.

The ongoing development of NeuroTrALE, supported by Lincoln Laboratory and the National Institutes of Health (NIH), aims to further enhance its machine learning capabilities. The software is being integrated with Google’s Neuroglancer, a web-based viewer for neuroscience data, enabling more dynamic data visualization and collaborative annotation. Adam Michaleas, a high-performance computing engineer at Lincoln Laboratory, notes that NeuroTrALE “offers a platform-agnostic solution that can be quickly deployed across various computing environments, making it accessible for researchers everywhere.”

The team’s ultimate goal is to make NeuroTrALE a fully open-source tool, aligning with NIH’s mission to share research products. Gjesteby emphasizes that “NeuroTrALE is part of a grassroots effort in the scientific community to share data and algorithms openly, accelerating the mapping of the human brain for research and drug development.”