Bridging Single Distortion Artifacts and Mmultifactorial Clinical Quality: Few-shot Biparametric MRI Quality Assessment via Distortion-trained Prototypical Networks

Abstract: Clinical prostate multi-parametric MRI relies heavily on high-quality diffusion-weighted imaging (DWI), yet reading DWI is frequently compromised by geometric distortion, often caused by rectal air. Assessing quality via the PI-QUAL scoring system is an emerging clinical standard, but it is subjective, time-consuming and suffers from a class imbalance where low-quality cases are diverse and relatively scarce. Using the PRIME clinical trial as an example, there are $6\%$ images with PI-QUAL scores lower than 4, $87\%$ of DWI issues are due to distortion. Many of the other clinical quality issues are under-represented. To address this common dual-scarcity of annotated clinical data, we propose a few-shot biparametric prototypical network for automated image quality assessment (IQA). Our framework utilizes a dual-branch 3D ResNet to fuse T2-weighted and DWI features, providing anatomical context to distinguish true morphology from distortion. To handle real-world heterogeneity, we introduce feature-wise linear modulation (FiLM) and a gradient reversal layer (GRL) to align feature distributions conditioned on varying b-values while suppressing acquisition-related biases. We demonstrate that a model meta-trained solely on comparatively objective, readily obtainable distortion labels can effectively adapt to predicting complex, multi-factorial clinical quality scores such as PI-QUAL using only five representative samples. Experimental results on two datasets show that our method significantly outperforms few-shot learning baselines for this challenging IQA task, offering a practically feasible and data-efficient solution for standardizing prostate MRI quality control in clinical workflows.

• The paper introduces a two-stage prototypical meta-learning framework that combines 3D DWI and T2WI features, masked normalization, b-value conditioning, and adversarial invariance for robust MRI quality assessment.
• The method transfers from distortion detection to multifactorial PI-QUAL classification on the PRIME dataset, reaching 72.85% balanced accuracy with only five support samples per class and outperforming few-shot baselines.
• The results show strong data efficiency and improved out-of-domain performance, but support-set selection, limited external validation, and the remaining classification errors mean the system should assist rather than replace radiologist scoring.

Motivation and problem setting

Automated image quality assessment (IQA) for prostate biparametric MRI is constrained by what the authors term a dual-scarcity problem. First, low-quality scans are rare even in optimized clinical trials: in the PRIME trial, only 6% of images receive PI-QUAL scores below 4, and 87% of DWI quality issues are attributable to geometric distortion, typically induced by rectal air during echo-planar acquisition. Second, the few available low-quality cases are dominated by a single artifact type, leaving other clinically relevant quality factors under-represented. PI-QUAL scoring itself is a multifactorial, subjective, radiologist-assigned measure with substantial inter-observer variation, and standard remedies are inadequate: resampling risks overfitting, conventional augmentation cannot simulate echo-planar geometric distortion, and k-space simulation requires auxiliary data such as field maps that retrospective cohorts rarely possess.

The central premise of the paper is that a model meta-trained on a comparatively objective and more abundant proxy task — binary DWI distortion labeling — can transfer to the subjective, multi-factorial PI-QUAL scoring task with only five labeled samples per class. This is a strong transfer claim, since a low PI-QUAL score does not exclusively imply poor DWI; distortion is merely the dominant, but not the sole, downgrade factor.

Method

The framework is a two-stage prototypical meta-learning pipeline. In the first stage, episodic $N$-way $K$-shot training ($N=2$, distorted vs. undistorted) is performed on a curated distortion dataset. A dual-branch 3D ResNet-18 encodes DWI and T2WI volumes in parallel; the T2WI branch supplies anatomical context so the model can distinguish true morphology from geometric displacement of the prostate in DWI. Multi-scale features from four ResNet blocks of both branches are combined with a CBAM-style attention fusion module, pooled, and mapped to embeddings from which class prototypes are computed by averaging L2-normalized support embeddings. Query predictions use temperature-scaled cosine similarity to prototypes with a cross-entropy prototypical loss.

Three components target real-world heterogeneity:

• Masked instance normalization (MiN): foreground voxels, selected by mean-intensity thresholding, are normalized to zero-mean unit-variance while background voxels are zeroed, preventing prototypes from encoding background intensity variation.
• Feature-wise linear modulation (FiLM): a one-hot b-value vector is embedded and linearly projected to per-channel scale and shift parameters applied to DWI-branch feature maps, conditioning representations on acquisition b-value.
• Gradient reversal layer (GRL): features are passed through a GRL to a b-value discriminator trained with cross-entropy; the reversed gradients ($-\alpha$ scaling) force b-value-invariant embeddings, with $\lambda=0.5$ weighting the adversarial loss.

In the second stage, all network parameters are frozen. Adaptation to a new dataset or task consists only of recomputing prototypes from $K=5$ representative support samples per class chosen to span acquisition conditions and morphology; the remaining samples form the query set.

Experimental setup

Meta-training uses a private dataset (Dataset 1) of 1027 high b-value ($\geq 1400$ s/mm$^2$) DWI volumes from 851 patients, each paired with T2WI and a prostate mask; distortion labels were obtained by manually assessing geometric consistency between the T2WI-based mask and the aligned DWI (790 undistorted, 237 distorted), split 8:2 at the patient level. Out-of-domain and cross-task evaluation uses PRIME (483 cases), with 28 radiologist-labeled distorted cases and 31 low-quality (PI-QUAL < 4) cases. Metrics are balanced accuracy (B-ACC), sensitivity at 80% specificity, and specificity at 80% sensitivity, averaged over 10 independent meta-training runs, with one-tailed Wilcoxon signed-rank significance tests against the best competing few-shot method. Baselines include Prototypical Networks, Siamese-Prototype Network, Semantic Guided Prototype Learning, Prototype-Neighbor Networks, a fully supervised upper bound (FSL), and a supervised model trained from scratch on only five support samples per class (FSL-5).

Results

Task / Dataset	Method	B-ACC (%)	Sen@80Spe (%)	Spe@80Sen (%)
Distortion, Dataset 1 (in-domain)	Ours	85.82 (1.31)	93.39* (1.03)	89.90 (1.51)
Distortion, PRIME (out-of-domain)	Ours	69.75 (5.44)	61.11 (5.91)	61.10 (5.94)
PI-QUAL, PRIME (cross-task)	Ours	72.85* (3.17)	65.22* (5.87)	66.11* (3.43)
PI-QUAL, PRIME (FSL-5 baseline)	Supervised, 5-shot	52.03 (5.96)	30.88 (2.51)	17.02 (2.00)

Asterisks denote statistical significance ($p<0.05$, one-tailed Wilcoxon signed-rank test) versus the best competing few-shot method.

In-domain performance is competitive across all prototypical methods, with the proposed model best overall and significantly better on Sen@80Spe, indicating improved detection of distorted DWI at fixed specificity. The fully supervised FSL upper bound reaches 88.63% B-ACC and 97.02% Sen@80Spe, so the few-shot model retains a measurable gap to full supervision in-domain — the method's advantage is data efficiency, not in-domain accuracy.

Out-of-domain generalization is where the contribution is most pronounced. Under domain shift and severe class imbalance on PRIME, the proposed method outperforms all baselines on all three metrics, with the strongest baselines (SPNet, SGPL) reaching only ~64–65% B-ACC and ~52–53% Sen@80Spe versus 69.75% and 61.11% for the proposed model. This indicates materially reduced risk of both missed detections and false rejections, which the authors argue is critical for clinical deployment.

Cross-task transfer to PI-QUAL is the headline claim: a model that never saw a PI-QUAL label achieves 72.85% B-ACC on PRIME using five support samples per class, outperforming all baselines with improvements of 10.15%–21.12% across metrics. The FSL-5 baseline collapses to near-chance behavior (17.02% Spe@80Sen), underscoring that direct supervised training on five imbalanced clinical samples is not viable. The result supports the paper's thesis that distortion-sensitive features transfer to the multifactorial clinical quality task, though it should be read as evidence of partial transfer: with 72.85% B-ACC, roughly a quarter of cases remain misclassified, so the model is a triage aid rather than a replacement for radiologist scoring.

Ablation via Grad-CAM visualizations and quantitative deltas confirms each component's contribution. Removing MiN causes attention leakage into rectal air and extra-prostatic tissue — a plausible shortcut given class imbalance. Removing the T2WI branch degrades distortion localization for lack of anatomical reference. Removing FiLM or GRL yields diffuse, low-intensity activations under varying b-values. Quantitatively, FiLM and GRL are the largest contributors, with average gains of 14.48%–19.64% on Sen@80Spe and Spe@80Sen, followed by the T2WI branch (10.63%–11.63%) and MiN (6.84%–9.17%).

Limitations and open questions

The paper is candid that the approach is inherently sensitive to support set selection; the authors mitigate this by manually choosing diverse samples that function as expert-defined reference templates, but this reintroduces a manual curation step that the automation pipeline does not remove, and no systematic study of support-set sensitivity (e.g., random versus curated selection) is reported. The distortion labels in Dataset 1 are themselves manually assessed against T2WI masks, so the "objective" proxy task retains subjective annotation. Evaluation is limited to two datasets from UK-centric clinical contexts, with only 28 distorted and 31 low-PI-QUAL cases in PRIME; the generalization of the transfer claim to other artifact types, scanners, and populations remains untested. The framework is demonstrated only for binary PI-QUAL thresholds (< 4 vs. ≥ 4); extending to finer-grained scoring, such as separating optimal (PI-QUAL 5) from sub-optimal scans, and to downstream effects on automated PI-RADS assessment, is left as open work.

Conclusion

This paper formulates prostate MRI IQA under a dual-scarcity regime and shows that distortion-supervised prototypical meta-learning, with FiLM-based b-value conditioning, adversarial b-value invariance, and T2WI-referenced anatomical context, transfers to multifactorial PI-QUAL assessment with five samples per class, achieving statistically significant gains over few-shot baselines (72.85% B-ACC on PRIME). The result substantiates the claim that physically grounded, objectively labeled artifact tasks can serve as effective pretexts for subjective clinical quality scoring, while leaving support-set robustness, broader artifact coverage, and fine-grained quality grading as open questions.