PERCEPT-Net: A Perceptual Loss Driven Framework for Reducing MRI Artifact Tissue Confusion

Abstract: Purpose: Existing deep learning-based MRI artifact correction models exhibit poor clinical generalization due to inherent artifact-tissue confusion, failing to discriminate artifacts from anatomical structures. To resolve this, we introduce PERCEPT-Net, a framework leveraging dedicated perceptual supervision for structure-preserving artifact suppression. Method: PERCEPT-Net utilizes a residual U-Net backbone integrated with a multi-scale recovery module and dual attention mechanisms to preserve anatomical context and salient features. The core mechanism, Motion Perceptual Loss (MPL), provides artifact-aware supervision by learning generalizable motion artifact representations. This logic directly guides the network to suppress artifacts while maintaining anatomical fidelity. Training utilized a hybrid dataset of real and simulated sequences, followed by prospective validation via objective metrics and expert radiologist assessments. Result: PERCEPT-Net outperformed state-of-the-art methods on clinical data. Ablation analysis established a direct causal link between MPL and performance; its omission caused a significant deterioration in structural consistency (p < 0.001) and tissue contrast (p < 0.001). Radiologist evaluations corroborated these objective metrics, scoring PERCEPT-Net significantly higher in global image quality (median 3 vs. 2, p < 0.001) and verifying the preservation of critical diagnostic structures. Conclusion: By integrating task-specific, artifact-aware perceptual learning, PERCEPT-Net suppresses motion artifacts in clinical MRI without compromising anatomical integrity. This framework improves clinical robustness and provides a verifiable mechanism to mitigate over-smoothing and structural degradation in medical image reconstruction.

• The paper introduces a Motion Perceptual Loss (MPL) that enhances artifact suppression by effectively distinguishing between true anatomical tissue and motion-induced artifacts.
• It integrates a residual U-Net with multi-scale recovery and dual attention modules, leading to significant gains in metrics like SSIM and CNR.
• Extensive experiments and radiologist assessments confirm that PERCEPT-Net reduces diagnostic errors and the need for repeat scans in clinical settings.

PERCEPT-Net: A Perceptual Loss-Driven Framework for MRI Motion Artifact Correction

Introduction and Motivation

The restoration of motion artifact-corrupted MRI sequences remains a principal obstacle for robust, high-fidelity neuroimaging. Clinical imaging is particularly susceptible to patient motion, resulting in artifacts such as blurring, ghosting, and distorted SNR profiles that confound the detection of subtle pathologies and compromise diagnostic utility. Conventional solutions—ranging from physical stabilization and k-space trajectory manipulation to post-hoc correction algorithms—present practical and physiological constraints, and physics-driven and data-driven deep learning approaches have demonstrated limitations in handling the heterogeneity of real-world artifacts. Critically, the inability of extant supervised models to disambiguate between genuine anatomical tissue and motion-induced artifacts results in over-smoothing, tissue loss, and artifact hallucinations, especially due to the inadequacies in using generic natural image perceptual losses for medical data.

Methodology

PERCEPT-Net operationalizes a new paradigm for artifact suppression by explicitly embedding perceptual supervision sensitive to MRI domain-specific artifact and anatomy differentiability. Its principal innovation, Motion Perceptual Loss (MPL), enables direct optimization of the discriminative boundary between genuine tissue and motion-induced artifact patterns.

Data Acquisition and Cohort Construction

A multi-center dataset was collated, comprising 12,110 axial 2D MR sequences from 664 patients, partitioned using stringent patient-level stratification into overlapping and non-overlapping training, validation, and external test splits. Training utilizes both real matched pairs (corrupted and re-scanned artifact-free images) and simulated artifact sequences generated via controlled k-space phase perturbation and subsampling to induce a realistic spectrum of artifact severity. For perceptual loss modeling, a balanced dataset of authentic motion-corrupted and authentic artifact-free unpaired sequences was assembled.

Network Architecture

The backbone is a residual U-Net with symmetrized encoder-decoder topology, enhanced by:

• Multi-scale recovery modules: Embedded at each decoding stage, aggregating multi-receptive-field features (3×3, 5×5, 7×7 convolutions) for both global and fine-grained structure recovery.
• Dual attention mechanisms: Combining squeeze-and-excitation channel attention (adaptive channel weighting) and spatial attention (highlighting salient anatomical regions) to preserve diagnostically salient tissues and localize artifact suppression.
• Motion Perceptual Loss (MPL): Unique to this framework, MPL is derived from a VGG19 network pre-trained on ImageNet and subsequently fine-tuned (with staged freezing) on MRI-specific artifact discrimination using a binary (artifact vs. anatomy) classification objective. Only high-level convolutional layers (Conv4 and Conv5 blocks) are adapted, ensuring the loss function is sensitive to mid-/high-level MRI-specific patterns. After fine-tuning, MPL is computed as the L1 distance between feature maps of cleaned and ground-truth sequences as extracted by Conv4_4 and Conv5_4, thus penalizing perceptual disagreement at intermediate feature levels relevant for MRI.

Training and Optimization

The objective function is a mix of pixel (L1), SSIM, k-space (focal frequency), adversarial, and MPL losses, with empirically tuned weights ensuring perceptual loss maintains predominant influence (e.g., $a_\text{L1}=0.3$, $a_\text{MPL}=0.25$). Training uses the Adam optimizer, cosine annealing scheduler, and careful patient-level separation to avoid data leakage.

Experimental Evaluation

Quantitative Metrics

Objective performance was quantified using PSNR, SSIM, CNR/SNR (tissue and signal contrast), LPIPS, and FID, controlling for multiple comparisons (FDR) and employing appropriate parametric and nonparametric hypothesis tests.

Ablation Studies

• Data authenticity: Training with authentic clinical data significantly (14.9% higher SSIM, 17.5% higher CNR) outperformed synthetic-only training, highlighting the clinical realism gap in simulated artifacts.
• Impact of MPL: Ablating MPL led to statistically significant reductions in SSIM and CNR (all $p<0.001$), with explicit demonstration that the MPL is causally essential for anatomical fidelity and artifact suppression.
• Integrated architecture: Multi-scale and attention modules contributed incrementally, but full integration (PERCEPT-Net) yielded the largest composite improvements (e.g., +6.3% SSIM, +39.7% CNR over multi-scale-attention-only variants).

Artifact Severity Stratification

PERCEPT-Net maintained robust improvement for moderate and severe artifacts, with T1 and T2 sequences showing SSIM increases from 0.609→0.754 (moderate) and 0.393→0.555 (severe), and analogous gains in CNR. Gains are more pronounced for more severe corruption, confirming that MPL confers resilience in challenging anatomical-ambiguous regions.

Radiologist Assessment

Blinded radiologist scoring on a 5-point Likert scale across seven critical anatomical regions confirmed significant (all $p<0.001$) improvement for PERCEPT-Net over corrupted input, with ICC inter-rater reliability of 0.800 for the restored images. The rate of diagnostically unacceptable images was reduced from 51.45% (raw) to 27.8% (restored), verifying direct clinical relevance.

Theoretical and Practical Implications

The core advancement is the MPL, which enforces feature-level separation between artifact and anatomy classes, operationalized with pathologically plausible clinical motion corruptions. This approach addresses the core pathology of artifact-tissue confusion, surpassing pixel- or SSIM-centric constraints. The architecture's multi-scale and attention mechanisms confer resilience against anatomical variability, while adversarial training supports distributional realism.

Practically, the framework supports robust deployment in multi-center clinical workflows, reduces unnecessary repeat scanning, and preserves subtle anatomical features critical for downstream tasks. The formation of artifact-aware perceptual losses sets a new methodological standard for future artifact suppression and image-to-image translation tasks in medical imaging, underlining the necessity of domain-specific perceptual objectives for clinical adoption.

Limitations and Future Directions

Systematic validation is constrained to a delimited range of MRI protocols and vendors (1.5T), with adversarial components increasing computational requirements. Future directions should include protocol/vendor diversity (e.g., 3T systems), adaptation to specialized sequences, cross-modality generalization, model compression for real-time deployment, and scaling to broader multi-center datasets. The framework's modular structure also allows exploration of alternative backbone architectures and loss function designs, potentially integrating diffusion-based or transformer mechanisms alongside MPL.

Conclusion

PERCEPT-Net, centered on a motion artifact-aware perceptual loss, provides a robust, clinically validated paradigm for MRI artifact correction. By resolving the artifact-tissue confusion through explicit feature discrimination, it achieves superior anatomical fidelity and artifact suppression compared to existing methods, with direct implications for clinical efficiency and diagnostic accuracy. Future research should further validate this approach across modalities and imaging environments, and continue developing domain-aligned perceptual supervision frameworks as the foundation for clinically trustworthy medical image restoration.

---

Reference: "PERCEPT-Net: A Perceptual Loss Driven Framework for Reducing MRI Artifact Tissue Confusion" (2604.10439)