Unsupervised Artifact Detection and Quantification via Contrastive Learning with Noise Reference

Levente Lippenszky, Hongxu Yang, Jose de Arcos, Laszlo Rusko

Keywords: MR artifact detection · Image quality control · Contrastive learning.

Abstract:

Artifacts in MR images can degrade diagnostic utility and
compromise the performance of downstream algorithms. Deep neural networks
are particularly sensitive to such artifacts and can produce inaccurate
or biased outputs. Automated artifact detection is therefore essential
for improving clinical efficiency and ensuring high-quality training data.
In this work, we propose a contrastive learning approach that structures
the embedding space to position images with higher artifact levels closer
to a noise reference. This enables unsupervised artifact detection and
quantification by computing the cosine similarity between the image and
noise embeddings at test time. Extensive experiments showed that our
method outperforms existing unsupervised approaches in detecting various
types of MR artifacts, including motion, ghosting, aliasing, metal and
gas, on prostate T2-weighted and brain T1-weighted images. In addition,
it achieved the highest performance in motion artifact quantification by
a substantial margin, highlighting its ability to learn rich representations
of artifact severity.

Submission Number: 2