Deep Learning Artifact Detection in Medical Imaging

A deep learning network is introduced as an intermediary component between the CT scanner and the reconstruction pipeline. The network receives sinogram data and reconstructed images as input, performs artifact detection and correction, and outputs corrected images. This intermediary approach allows the system to leverage pre-trained deep learning models without fundamentally redesigning the entire imaging system, thus improving image quality while managing complexity through modular integration.

The deep learning network is pre-trained on large datasets of annotated medical images with various artifacts before deployment. During actual imaging workflows, the pre-trained network performs rapid inference to detect and correct artifacts in real-time. This preliminary training phase separates the computationally intensive model development from the clinical workflow, allowing automated artifact correction to run efficiently during patient scans without adding significant computational burden to the imaging workflow.

The artifact detection and correction functionality is extracted as a separate deep learning network module that operates independently from the main CT reconstruction pipeline. The network processes only the necessary data (sinograms and reconstructed images) to identify and correct artifacts, rather than reprocessing the entire imaging workflow. This extraction allows for optimized computational resource allocation and energy-efficient processing of artifact-prone regions.

The deep learning network is designed to autonomously detect artifacts, determine the appropriate correction strategy, and apply corrections without requiring manual intervention from technologists or radiologists. The system self-manages the entire artifact correction workflow, from detection to correction to validation, thereby eliminating the need for rescans while maintaining simplicity in operational procedures despite the underlying computational complexity.