[Paper Review] Motion Artifacts Detection in Short-scan Dental CBCT Reconstructions

Abdul Salam Rasmi Asraf Ali, Andrea Fusiello+3 more|arXiv (Cornell University)|Apr 20, 2023

Medical Imaging Techniques and ApplicationsMedicine3 citations

TL;DR

This paper proposes a deep learning framework to detect motion artifacts in short-scan dental CBCT reconstructions using simulated motion data and data augmentation. By classifying projection subsets as motion-free or artifact-affected, the method enables artifact-free short-scan volume reconstruction without correction algorithms, achieving high AUC-PR (up to 0.983) on simulated data and robust performance (AUC-PR = 1.000) on real motion-affected scans, demonstrating feasibility for clinical use.

ABSTRACT

Cone Beam Computed Tomography (CBCT) is widely used in dentistry for diagnostics and treatment planning. CBCT Imaging has a long acquisition time and consequently, the patient is likely to move. This motion causes significant artifacts in the reconstructed data which may lead to misdiagnosis. Existing motion correction algorithms only address this issue partially, struggling with inconsistencies due to truncation, accuracy, and execution speed. On the other hand, a short-scan reconstruction using a subset of motion-free projections with appropriate weighting methods can have a sufficient clinical image quality for most diagnostic purposes. Therefore, a framework is used in this study to extract the motion-free part of the scanned projections with which a clean short-scan volume can be reconstructed without using correction algorithms. Motion artifacts are detected using deep learning with a slice-based prediction scheme followed by volume averaging to get the final result. A realistic motion simulation strategy and data augmentation has been implemented to address data scarcity. The framework has been validated by testing it with real motion-affected data while the model was trained only with simulated motion data. This shows the feasibility to apply the proposed framework to a broad variety of motion cases for further research.

Motivation & Objective

• To address the challenge of motion artifacts in dental CBCT, which degrade diagnostic accuracy and are common in clinical practice.
• To develop a framework that detects motion-free projection subsets for short-scan reconstruction, avoiding the need for complex correction algorithms.
• To overcome data scarcity in motion-affected CBCT by using realistic motion simulation and augmentation for model training.
• To validate the framework’s robustness on real-world motion-affected scans despite being trained only on simulated data.
• To enable clinically viable, artifact-free CBCT reconstructions using only projection classification and volume averaging.

Proposed method

• A slice-based deep learning framework classifies short-scan volumes reconstructed from selected projection subsets as positive (motion-affected) or negative (motion-free).
• Realistic motion simulation generates synthetic motion data with various types (nodding, tilting, lateral rotation, tremor) and varying intensities.
• Data augmentation techniques are applied to increase diversity and generalization of the simulated motion data.
• Four short-scan volumes are reconstructed from different projection subsets to cover 360°, and results are averaged to produce a final diagnostic-quality volume.
• EfficientNet V2 is selected as the backbone model due to its superior performance and generalization to real-world motion.
• Volume averaging of multiple short-scan reconstructions reduces residual artifacts and improves diagnostic consistency.

Experimental results

Research questions

• RQ1 Can a deep learning model trained exclusively on simulated motion data detect motion artifacts in real motion-affected dental CBCT scans with high accuracy?
• RQ2 How effective is realistic motion simulation and data augmentation in mitigating data scarcity for training motion artifact detection in CBCT?
• RQ3 Can short-scan reconstructions from motion-free projection subsets produce diagnostically sufficient image quality without correction algorithms?
• RQ4 Which deep learning architecture generalizes best across diverse motion types and real-world scan variations?
• RQ5 Can the framework distinguish motion artifacts from other non-motion-related artifacts such as beam hardening or photon starvation?

Key findings

• The framework achieved an AUC-PR of 0.983 on simulated motion data using EfficientNet V2, outperforming other architectures.
• On real motion-affected scans, the model achieved a perfect AUC-PR of 1.000 with EfficientNet V2, demonstrating strong robustness to real-world motion.
• EfficientNet V2 and EfficientNet showed superior generalization to real data compared to ResNet and DenseNet, indicating better domain generalization.
• The framework successfully discriminated motion artifacts from non-motion-related artifacts such as beam hardening and noise in phantom scans.
• Volume averaging of multiple short-scan reconstructions significantly improved image consistency and reduced residual artifacts.
• The model’s ability to detect motion in real scans despite being trained only on simulated data validates the effectiveness of the simulation strategy for future research.