RPI ID: 2018-021-201 / 2018-021-601
Innovation Summary:
This invention introduces a method for training convolutional neural networks (CNNs) to reduce artifacts in CT images using pseudo ground truth data. The system includes an estimated ground truth generator that produces reference images from artifact-laden scans, enabling the CNN to learn artifact correction without requiring ideal, artifact-free datasets. The approach supports both supervised and semi-supervised learning and is adaptable to various artifact types and imaging protocols. This innovation significantly improves the practicality and scalability of AI-based CT enhancement in clinical environments.
Challenges / Opportunities:
Obtaining clean, high-quality CT images for supervised AI training is difficult, especially in real-world clinical settings. Traditional methods depend on manually curated or simulated datasets, which are often limited in scope. This invention addresses those limitations by generating pseudo ground truth images that mimic real-world conditions, allowing CNNs to be trained effectively even when true ground truth is unavailable. It opens opportunities for scalable, AI-driven artifact reduction across diverse imaging environments.
Key Benefits / Advantages:
✔ Enables CNN training without perfect reference data
✔ Effectively reduces artifacts and improves image clarity
✔ Adaptable to various artifact types and imaging protocols
✔ Supports supervised and semi-supervised learning
✔ Reduces data collection and labeling costs
✔ Compatible with existing CT systems and AI workflows
Applications:
• Medical imaging and diagnostics
• AI-assisted radiology
• Low-dose and emergency CT scans
• Imaging research and development
• Clinical decision support systems
Keywords:
#CTartifactreduction #CNNtraining #pseudogroundtruth #deeplearning #medicalimaging #AIradiology
Intellectual Property:
US Issued Patents 11,120,551 B2 and 11,727,569 B2
