Data Augmentation Reduces Motion Artifacts Impact on MRI Lower Limb AI Segmentation

Overview

Motion artifacts significantly degrade MRI image quality and reduce the accuracy of AI models for lower limb segmentation and torsion quantification. This study demonstrates that domain-specific data augmentation during AI training stabilizes model performance across varying artifact severities, improving robustness in clinical scenarios.

Background

MRI is a preferred imaging modality due to excellent soft tissue contrast and lack of ionizing radiation but is highly susceptible to motion artifacts, affecting up to one-third of clinical sequences. These artifacts impair diagnostic accuracy and increase scan repetition rates, impacting patient comfort and healthcare costs. AI models for musculoskeletal imaging, particularly for bone segmentation and alignment quantification, are sensitive to image quality degradation. Data augmentation techniques may enhance AI robustness by increasing training data diversity and quality, yet their effect on artifact mitigation remains underexplored.

Data Highlights

The study used axial T2-weighted 2D turbo-spin echo MRI sequences from 93 patients for training and validation, totaling 186 lower limbs with expert manual segmentations. A U-net CNN was trained with three different data augmentation strategies to segment femur, tibia, and fibula and quantify torsional alignment. A prospectively acquired test set of 20 healthy participants was used to evaluate model performance across increasing motion artifact severity.

Key Findings

• Motion artifacts cause progressive degradation in AI segmentation accuracy and torsion measurement precision.
• Models trained with domain-specific data augmentation showed significantly improved robustness to motion artifacts compared to models trained without augmentation.
• Data augmentation strategies enhanced the diversity and quality of training data, stabilizing AI performance across artifact severities.
• Automated landmark identification and torsion quantification remained reliable when augmented training was applied, reducing intra- and inter-reader variability.
• Prospective test data confirmed that augmentation-trained models maintained clinical utility despite artifact presence.

Clinical Implications

Incorporating domain-specific data augmentation into AI model training can mitigate the negative impact of motion artifacts on MRI-based lower limb segmentation and alignment quantification. This approach may reduce the need for scan repetitions and improve diagnostic confidence. Clinicians and developers should consider augmentation strategies to enhance AI robustness in routine MRI workflows.

Conclusion

Data augmentation is a viable strategy to enhance AI model resilience against motion artifacts in MRI, supporting more reliable automated musculoskeletal assessments. This advancement may facilitate broader clinical adoption of AI tools despite inherent imaging challenges.

References

1. Schock et al 2023 -- AI Model for Lower Limb Alignment Quantification
2. Lee et al -- Method for Reference Line Definition in Torsion Measurement