By: Isabelle Banaie and Kimberly Mann Bruch
Researchers from the University of California San Diego have introduced a new automated approach that makes brain analysis faster, more accurate and easier to understand. Their work was published in the AI Open journal.
David Kleinfeld, distinguished professor in the School of Physical Sciences Physics Department, worked on the project with several colleagues including Yoav Freund, a professor in the Jacobs School of Engineering Computer Science and Engineering Department and an affiliate with the School of Computing, Information and Data Sciences Halıcıoğlu Data Science Institute (HDSI). The team developed a computer system that can identify specific brain regions based on the shapes and distribution of cells, rather than just pixel intensity or black-box algorithms that are difficult to interpret.
“Our approach marks a significant step towards explainable brain analysis, allowing researchers to see precisely which features influence decisions,” Kleinfeld said. “Instead of using unclear neural networks that can be a black box, our method relies on interpretable features derived from cell shapes.”
He said that this explainability provides scientists with the ability to identify cases where the classification is not confident, which are either discarded or corrected. This is crucial for scientific validity.
The core of this new system involves first detecting individual cells in high-resolution images, then analyzing their shapes through advanced mathematical techniques like diffusion mapping to reduce complexity while preserving meaningful biological information. These cell features are combined to describe regions of the brain, which the system classifies using a machine learning method called XGBoost, which can indicate the ways that features such as cell orientation or size most influence decisions, making the process explainable.
The researchers compared their method to prior techniques that used neural networks that were harder to interpret, demonstrating comparable accuracy but with the added benefit of clarity. Their approach also remains effective across different tissue staining and imaging techniques, making it versatile.
“In one example, we generated probability maps for brain structures that closely match what a trained anatomist would identify under a microscope,” Freund said. “What’s exciting is developing a workflow which speeds up the work of the anatomist while maintaining the scientific validity of the results.”
“By enabling an open, interpretable analysis process, our research paves the way for faster, more reliable brain mapping, with applications in studying brain circuitry, development and disease,” Kleinfeld said. “Our goal is to merge computational power with biological understanding, so researchers can trust what the machine learns and use it as a partner in discovery.”
The work was funded by the U.S. National Institutes of Health BRAIN Initiative (grant nos. U19 NS107466 and U19 NS137920).
