Plus.DLR

Patient-Friendly Imaging with Plus.DLR

Solving the Challenges Faced by Physicians: The Medical Imaging Noise Problem

　・ Noise in CT and MRI images occurs due to equipment performance and imaging conditions.
　・ Noise increases the risk of missing lesions or misdiagnosis.
　・ There is an inherent trade-off between radiation dose and image quality.

Core Functions:　

　・ AI Reduces Noise and Enhances Image Clarity
　・ Technology that compensates for image degradation in low-dose imaging
　・ Processing time does not interfere with the reading workflow
　 　 (approximately 2–3 minutes to process a CT image series of about 200 slices)

The Innovation of Plus.DLR:

AI Denoising Without Pretraining - Overcomes The Challenges of Traditional Deep Learning Models

The disadvantages of deep learning models come from pretraining

　・ Typically, deep learning-based noise removal requires pretraining.
　・ However, the scope of application is restricted by the items of pretraining data used (scanner, part of the body, dose, reconstruction
　　 function).
　・ Since some structures are memorized within the deep learning model parameters, structures that do not actually exist may be generated, 　　 while existing structures may be removed.

A New Approach: Noise2Noise[1], Eliminating the Need for Pretraining

　・ Learns in real-time*, making pretraining unnecessary.
　・ Not restricted by training dataset parameters – a single model can be applied to a wide range of medical images.
　・ The generation or removal of structures caused by pretraining is less likely to occur. Additionally, the network is designed to prevent the
　　 creation of artificial structures.

　 [1] https://arxiv.org/abs/1803.04189

　＊No human input (supervisory signal) is required, and the learning process is deterministic. Therefore, this is not classified as post-learning functionality.