[Paper Review] Automated SSIM Regression for Detection and Quantification of Motion Artefacts in Brain MR Images

Alessandro Sciarra, Soumick Chatterjee+5 more|arXiv (Cornell University)|Jun 14, 2022

Radiomics and Machine Learning in Medical Imaging4 citations

TL;DR

This paper proposes an automated, reference-free image quality assessment method for brain MR images using deep learning to regress the Structural Similarity Index (SSIM) and detect motion artefacts. By training ResNet-18 with contrast augmentation on artificially corrupted images, the model achieves high-precision SSIM prediction (μ = −0.0009, σ = 0.0139) and classification accuracy up to 97% for three motion severity classes.

ABSTRACT

Motion artefacts in magnetic resonance brain images can have a strong impact on diagnostic confidence. The assessment of MR image quality is fundamental before proceeding with the clinical diagnosis. Motion artefacts can alter the delineation of structures such as the brain, lesions or tumours and may require a repeat scan. Otherwise, an inaccurate (e.g. correct pathology but wrong severity) or incorrect diagnosis (e.g. wrong pathology) may occur. " extit{Image quality assessment}" as a fast, automated step right after scanning can assist in deciding if the acquired images are diagnostically sufficient. An automated image quality assessment based on the structural similarity index (SSIM) regression through a residual neural network is proposed in this work. Additionally, a classification into different groups - by subdividing with SSIM ranges - is evaluated. Importantly, this method predicts SSIM values of an input image in the absence of a reference ground truth image. The networks were able to detect motion artefacts, and the best performance for the regression and classification task has always been achieved with ResNet-18 with contrast augmentation. The mean and standard deviation of residuals' distribution were $μ=-0.0009$ and $σ=0.0139$, respectively. Whilst for the classification task in 3, 5 and 10 classes, the best accuracies were 97, 95 and 89\%, respectively. The results show that the proposed method could be a tool for supporting neuro-radiologists and radiographers in evaluating image quality quickly.

Motivation & Objective

• To develop an automated, reference-free image quality assessment (IQA) tool for brain MR images that detects and quantifies motion artefacts without requiring a ground-truth artefact-free image.
• To enable robust performance across diverse MR image contrasts (T1, T2, PD, FLAIR), resolutions, and orientations.
• To improve generalization by introducing contrast augmentation during training to simulate variability in clinical imaging protocols.
• To evaluate the method’s performance in both regression (SSIM prediction) and classification (motion severity levels) tasks.
• To provide a fast, objective, and reliable alternative to subjective radiological assessment of image quality.

Proposed method

• The method employs residual neural networks (ResNet-18 and ResNet-101) to regress SSIM values from single input MR images without requiring a reference image.
• Artificial motion artefacts were generated using two methods: one based on TorchIO’s implementation of Shaw et al.’s algorithm and a second custom-developed algorithm.
• Ground-truth SSIM values for training were computed by comparing artificially corrupted images with their original, artefact-free counterparts.
• Contrast augmentation was applied during training to simulate variations in image intensity and improve robustness across different imaging protocols.
• The model was trained in a fully supervised manner using SSIM values as regression targets, enabling prediction of image quality degradation levels.
• Classification into motion severity levels was performed by binning predicted SSIM values into predefined ranges (3, 5, or 10 classes).

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict the SSIM value of a motion-corrupted MR image without access to a reference artefact-free image?
• RQ2 How does contrast augmentation affect the robustness and performance of SSIM regression and classification models across diverse MR image contrasts and intensities?
• RQ3 What is the achievable accuracy of classifying motion artefact severity levels using predicted SSIM values from a reference-free model?
• RQ4 How does the proposed method compare to existing IQA tools like MRIQC in terms of speed, accuracy, and compatibility with clinical workflows?
• RQ5 To what extent can the model generalize to real-world clinical scans with varying contrasts, resolutions, and orientations?

Key findings

• The ResNet-18 model with contrast augmentation achieved the best regression performance, with a residual mean of −0.0009 and standard deviation of 0.0139, indicating near-perfect SSIM prediction accuracy.
• For motion classification, the model achieved 97% accuracy with three severity classes, 95% with five classes, and 89% with ten classes.
• The method demonstrated high robustness across different MR contrasts (T1, T2, PD, FLAIR), resolutions, and image orientations, as intended in the design.
• The inference time was significantly faster than MRIQC—8.84 seconds on GPU for 36 volumes versus 15 minutes on CPU for 12 volumes.
• The model required only 4GB RAM and ~1GB VRAM, making it suitable for deployment on standard clinical hardware, unlike MRIQC which requires 49GB RAM and multiple complex dependencies.
• The method showed 75% agreement with MRIQC on quality metrics, and 75% agreement with subjective expert assessments, validating its clinical relevance.