RACE-Net: A Recurrent Neural Network for Biomedical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Biomedical Engineering
• Computer Vision

Background:

• Level set based deformable models (LDM) require handcrafted velocities for medical image segmentation.
• Convolutional Neural Networks (CNNs) offer end-to-end learning but demand extensive data and computational resources, and struggle with boundary shape preservation.

Purpose of the Study:

• To introduce RACE-net, a Recurrent Neural Network-based solution to improve medical image segmentation.
• To address limitations of LDMs and CNNs by learning curve evolution velocities efficiently.

Main Methods:

• RACE-net models a generalized LDM with constant and mean curvature velocity.
• Curve evolution velocities are approximated using a feed-forward architecture inspired by multiscale image pyramids.
• The network learns velocities end-to-end, minimizing parameters, computation, and memory requirements.

Main Results:

• RACE-net achieved high Dice values (0.87-0.97) across three diverse segmentation tasks: optic disc/cup, cell nuclei, and left atrium.
• The model demonstrated utility as a generic, off-the-shelf architecture for biomedical segmentation.
• Validation was performed on public datasets for optic disc and cup in color fundus images, cell nuclei in histopathological images, and left atrium in cardiac MRI volumes.

Conclusions:

• RACE-net offers an efficient and effective solution for medical image segmentation.
• The Recurrent Neural Network approach overcomes limitations of traditional methods and CNNs.
• RACE-net provides a versatile and computationally efficient tool for various biomedical segmentation applications.