GLIMPSE: Generalized Locality for Scalable and Robust CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning, particularly convolutional neural networks (CNNs), is state-of-the-art for medical tomographic imaging.
• Current CNN approaches often overfit large structures and generalize poorly to out-of-distribution samples.
• High memory and computational demands of multiscale CNNs limit their use in realistic clinical resolutions.

Purpose of the Study:

• Introduce Glimpse, a local coordinate-based neural network for computed tomography.
• Address limitations of existing deep learning methods in medical imaging, including generalization and resource efficiency.
• Enable high-resolution medical image reconstruction with reduced computational complexity.

Main Methods:

• Developed Glimpse, a local coordinate-based neural network that reconstructs pixel values using neighborhood-associated measurements.
• Implemented a novel approach processing local measurement data for each pixel.
• Designed Glimpse to be fully differentiable for integration into existing deep learning architectures.

Main Results:

• Glimpse significantly outperforms CNNs on out-of-distribution samples in computed tomography.
• Achieved comparable or better performance than CNNs on in-distribution data.
• Maintained a memory footprint nearly independent of image resolution, enabling training on high-resolution images (e.g., 1024x1024) with 5GB memory.

Conclusions:

• Glimpse offers superior generalization and efficiency for medical tomographic imaging compared to traditional CNNs.
• The local reconstruction approach makes Glimpse suitable for high-resolution clinical applications.
• Glimpse's differentiability allows for advanced applications like correcting projection miscalibrations.