New paper by Devi Klein: Greater benefits of AI for finding small signals in 3D volumetric medical images

Modern medical imaging, such as digital breast tomosynthesis (DBT), produces three-dimensional (3D) images that allow doctors to look through layers of tissue to find signs of disease, such as tumors or microcalcifications. While these 3D scans give a clearer view than traditional two-dimensional (2D) mammograms, they also require radiologists to search through hundreds of image slices — a time-consuming process that increases the chance of missing small, subtle abnormalities that might only appear in a few slices.

In a new paper in the Journal of Medical Imaging, Vision & Image Understanding Lab grad student Devi Klein and colleagues at UCSB explored how a deep learning–based computer-aided detection (CADe) system could assist human observers when searching through these complex 3D images. An artificial intelligence (AI) model called a convolutional neural network (CNN) was used to highlight potentially suspicious areas with small boxes on the images. Sixteen trained participants were asked to find both small “microcalcification” signals and larger “mass” signals in simulated 2D and 3D breast images, both with and without the AI’s help, while their eye movements were tracked.

The findings showed that the AI assistant provided the biggest boost to search accuracy when participants searched 3D images for small target, such as microcalcifications, which are easily overlooked when scanning through large image stacks. The AI helped participants detect more of these tiny features, especially those observers who tended to move their eyes less and rely more on peripheral vision. For larger signals, like masses, the AI offered smaller improvements — suggesting that the system is most valuable when the task requires finding subtle features that are easily missed in 3D volumes

In simple terms, this research demonstrates when and why AI can help radiologists — helping them navigate complex 3D image data and reducing the risk of missing small but important details. The work highlights the growing potential of AI to enhance human performance in medicine, particularly in areas like breast cancer screening where early detection is critical. Beyond radiology, the study also provides insight into how human visual search and attention interact with AI guidance systems in real-world decision-making tasks.