Researchers at the Moscow Institute of Physics and Technology (MIPT) have developed a hybrid neural network architecture for left atrium segmentation in MRI scans. The model, named cerebellar-inspired, mimics the human cerebellum to improve stability against noise and artifacts.
The architecture features a recurrent block for data processing and a cerebellar module that predicts hidden image features. This allows the system to provide corrective signals, stabilizing computations when processing distorted clinical data.
During tests on the ATRIA dataset of 100 patients, the model increased the Dice coefficient accuracy from 0.801 to 0.815 on noisy data. It outperformed standard recurrent models in boundary precision, which is vital for planning catheter ablation.
The model demonstrates high computational efficiency, processing a 3D MRI volume in 18 milliseconds, compared to 23 seconds for the nnU-Net algorithm. It also requires 109 MB of video memory, which is 15 times less than the 1.5 GB used by nnU-Net.
The source code has been released on GitHub for public use. Future research will focus on testing the model against real-world clinical artifacts, such as patient breathing distortions, and integrating mechanisms that simulate inhibitory interneurons.
