PYRO-NN

PYRO-NN integrates computed tomography (CT) reconstruction algorithms into the TensorFlow deep learning framework to embed known operators as differentiable layers for medical image reconstruction.

Key Features:

• TensorFlow integration: Implements CT reconstruction operators as layers within TensorFlow, enabling embedding of known operators directly into neural networks.
• CUDA acceleration: Provides CUDA-accelerated implementations of parallel-, fan-, and cone-beam projectors and back-projectors to increase computational throughput.
• Reconstruction algorithms: Includes filtered back projection (FBP) and iterative reconstruction algorithms for processing data from real CT systems.
• Mathematical precision: Embeds mathematically precise known operators to maintain unambiguous solvability of the implemented operations.
• Geometry handling: Encodes geometry setup for projections and reconstructions fully integrated into TensorFlow.

Scientific Applications:

• Medical image reconstruction: Supports development and evaluation of deep-learning-based CT reconstruction methods to improve imaging quality and diagnostic accuracy.
• Hybrid reconstruction research: Enables experiments that combine traditional CT reconstruction methods with modern machine learning by embedding reconstruction operators as network layers.

Methodology:

Sets up geometry for projections and reconstructions within TensorFlow; implements CUDA-accelerated parallel-, fan-, and cone-beam projectors and back-projectors as TensorFlow layers; provides filtered back projection (FBP) and iterative reconstruction algorithms; and embeds known operators directly into neural networks to ensure mathematically unambiguous operations.