Overview

• The MTANN framework trains on just 68 CT scan cases to achieve an AUC of 0.92, outperforming Vision Transformer and 3D ResNet benchmarks.
• Training on a commercial laptop takes only 8 minutes and 20 seconds and generates diagnostic predictions in 47 milliseconds per patient case.
• By eliminating reliance on GPUs and large datasets, the model could extend diagnostic capabilities to rural clinics and improve detection of rare diseases.
• Its low computational footprint translates into substantially reduced energy consumption compared with conventional AI systems.
• Suzuki’s team received the Cum Laude Award at RSNA 2024 in recognition of its innovation and the Institute of Science Tokyo fosters the interdisciplinary research environment behind the breakthrough.