Radiation-Preserving Selective Imaging for Pediatric Hip Dysplasia: A Cross-Modal Approach

Conference: AAAI 2026 arXiv: 2511.18457 Code: None Area: Medical Imaging / Cross-Modal Learning / Selective Imaging Keywords: Developmental dysplasia of the hip, ultrasound–X-ray cross-modal learning, conformal prediction, selective imaging, self-supervised learning

TL;DR

This paper proposes an "ultrasound-first, radiation-preserving" cross-modal selective imaging strategy. By combining a self-supervised pretrained frozen encoder, a measurement-faithful lightweight head network, and a conformal-prediction-calibrated one-sided lower bound, the framework provides principled decisions on when ultrasound alone suffices and when additional X-ray imaging is warranted for diagnosing developmental dysplasia of the hip (DDH).

Background & Motivation

Developmental dysplasia of the hip (DDH) is a common pediatric orthopedic condition encompassing a spectrum of pathologies ranging from insufficient acetabular coverage to complete hip dislocation. Two complementary imaging modalities are used clinically:

Ultrasound (US): Used for early screening and management, reporting Graf α/β angles and femoral head coverage. Advantages include no ionizing radiation and broad accessibility.

X-ray (XR): More appropriate for evaluating acetabular development and surgical planning after ossification has progressed, measuring the acetabular index (AI), center-edge angle (CE), and IHDI grade. The key disadvantage is exposure to ionizing radiation.

Key Challenge: In pediatric care, minimizing ionizing radiation exposure is a fundamental principle, yet relying solely on ultrasound risks missing abnormalities detectable only on X-ray. Clinically, a clearly normal ultrasound (high α angle, adequate coverage) generally implies low marginal diagnostic value from X-ray; an abnormal or borderline ultrasound raises the value of proceeding to X-ray.

Limitations of Prior Work: - Automated ultrasound and automated X-ray analysis have each advanced independently (automatic standard plane detection, Graf measurement prediction, AI/CE angle estimation, etc.), but cross-modal US–XR learning remains extremely scarce. - No practical, measurement-grounded strategy exists for trading off radiation risk against diagnostic gain with explicit statistical guarantees. - Existing methods are largely black-box classifiers, lacking direct alignment with clinical decision thresholds.

The paper's motivation is to address this gap: constructing a calibratable, interpretable, and tunable "ultrasound-first" strategy that uses conformal prediction to provide finite-sample, distribution-free coverage guarantees, enabling clinical teams to make transparent trade-offs between radiation exposure and the risk of missed findings.

Method

Overall Architecture

The paper proposes a four-stage pipeline:

1. Self-supervised pretraining: Modality-specific ResNet-18 encoders are pretrained independently with SimSiam on large-scale unannotated registry data (37,186 ultrasound images + 19,546 X-ray images), then frozen.
2. Measurement-faithful head networks: Lightweight MLP heads are trained on top of the frozen encoders to directly predict clinically used named measurements (α/β angles, coverage, AI, CE, IHDI, etc.).
3. Conformal calibration: An affine bias correction is fitted on a calibration set, and one-sided conformal residual quantiles are computed, providing finite-sample marginal coverage guarantees for ultrasound predictions.
4. Selective imaging strategy: Calibrated ultrasound lower bounds are compared against clinical thresholds to decide "ultrasound only" versus "refer to X-ray," with decision curve analysis used to trade off radiation cost against missed-finding penalty.

Key Designs

1. SimSiam self-supervised pretraining: An encoder \(f_\phi\) is trained independently for each modality. Let the projection MLP be \(h_\phi\) and the prediction MLP be \(q_\phi\). Two augmented views \(v_1, v_2\) are generated from image \(x\), and the loss is the stop-gradient negative cosine similarity: $\(\mathcal{L}_{SSL} = -\frac{1}{2}\left[\frac{\langle p_1, \text{sg}(z_2)\rangle}{\|p_1\| \|\text{sg}(z_2)\|} + \frac{\langle p_2, \text{sg}(z_1)\rangle}{\|p_2\| \|\text{sg}(z_1)\|}\right]\)$ After 10 epochs of training, the projection and prediction heads are discarded and the encoder is frozen. Design Motivation: To leverage large volumes of unannotated imaging data for learning general visual representations, compensating for the scarcity of annotated samples.

2. Measurement-faithful head networks: A 512-dimensional global average-pooled feature \(u = f_\phi(x)\) is extracted from the frozen encoder, and a small MLP is trained on top:

3. Ultrasound head: Single hidden layer (128 units), 3 outputs predicting \(\hat{\alpha}, \hat{\beta}, \widehat{\text{cov}}\), with loss \(\mathcal{L}_{US} = \lambda_\alpha |\hat{\alpha} - y^\alpha| + \lambda_\beta |\hat{\beta} - y^\beta| + \lambda_{\text{cov}} |\widehat{\text{cov}} - y^{\text{cov}}|\)

4. X-ray head: Predicts AI and CE angles (MAE loss) with optional IHDI classification (cross-entropy)

The core design philosophy is measurement faithfulness — outputs are the named measurements used in daily clinical practice rather than black-box classifications, making the decision process fully traceable.

1. Conformal-calibrated one-sided lower bounds: Constructed in two steps:

2. (i) Affine bias correction: \(\tilde{y}^t = a_t \hat{y}^t + b_t\) is fitted on the calibration set to reduce systematic bias without retraining the head.

3. (ii) One-sided residual quantile: Residuals \(r_i^t = y_i^t - \tilde{y}_i^t\) are computed, and a conformal radius \(q_t^+(\delta_t)\) is calculated at miscoverage level \(\delta_t\), ensuring that under exchangeability \(y^t \geq \text{LB}_t(x; \delta_t)\) holds with probability \(\geq 1 - \delta_t\). The calibrated lower bound is \(\text{LB}_t(x; \delta_t) = \tilde{y}^t(x) - q_t^+(\delta_t)\).

4. Three selective imaging rules: Based on comparing calibrated lower bounds against clinical thresholds \(T_\alpha = 60°\) and \(T_{\text{cov}} = 50\%\):

5. Alpha-only: \(d_\alpha(x) = \mathbb{I}[\text{LB}_\alpha(x) \geq 60°]\)

6. Alpha OR Coverage: Either lower bound exceeding its threshold is sufficient to proceed with ultrasound alone
7. Alpha AND Coverage: Both lower bounds must exceed their thresholds to proceed with ultrasound alone

Sweeping over a grid of \((\delta_\alpha, \delta_{\text{cov}})\) values generates a family of policies. Conservative settings (small \(\delta\)) provide high coverage but low ultrasound pass rates; lenient settings (large \(\delta\)) increase the pass rate at the cost of higher missed-finding risk.

Loss & Training

• Self-supervised stage: SimSiam negative cosine similarity loss, 10 epochs per modality
• Supervised head networks: MAE loss for ultrasound (equal weights \(\lambda_\alpha = \lambda_\beta = \lambda_{\text{cov}} = 1\)), MAE + cross-entropy for X-ray
• Head networks are trained exclusively on the post-train set (30 subjects, 136 images); the encoder remains frozen throughout
• The calibration set (7 subjects, 28 images) is used only for bias correction and conformal calibration, not for training
• Data splits are strictly performed at the subject level to prevent leakage

Key Experimental Results

Main Results

Dataset: 75 paired subjects and 321 images drawn from a large registry. The evaluation set consists of 38 subjects (157 images), yielding N=77 strictly matched hip pairs after pairing.

Modality	Measurement	MAE	Notes
Ultrasound	α angle	9.69°	Frozen encoder + lightweight head
Ultrasound	β angle	11.25°	Same
Ultrasound	Femoral head coverage	13.97 pp	Percentage points
X-ray	Acetabular index (AI)	7.60°	Frozen encoder + lightweight head
X-ray	Center-edge angle (CE)	8.93°	Same

Selective imaging strategy results (N=77 strictly matched pairs):

Rule	Miscoverage \(\delta_\alpha / \delta_{\text{cov}}\)	US Pass Rate	X-ray Usage Rate
AND	0.10 / 0.10	0.00	1.00
AND	0.20 / 0.20	0.00	1.00
OR	0.35 / 0.35	0.43	0.57
OR	0.40 / 0.40	0.55	0.45

Ablation Study

Configuration	Key Result	Notes
Conservative AND 0.10/0.10	Coverage ~0.90 (α), ~0.94 (cov)	Safe but nearly all cases referred to X-ray
Lenient OR 0.40/0.40	US pass rate 55%	Substantially reduces X-ray use but increases missed-finding risk
Calibration radius (δ=0.10)	\(q_\alpha^+ = 10.75°\), \(q_{\text{cov}}^+ = 28.74\) pp	Conservative lower bounds
Decision curve analysis	OR strategy optimal at high radiation cost λ	Explicit radiation vs. safety trade-off

Key Findings

1. Frozen encoder + lightweight head is competitive: A small MLP head trained on only 30 subjects atop a frozen self-supervised encoder achieves measurement accuracy comparable to single-modality methods using larger networks.
2. Conformal coverage follows theoretical expectations: Empirical coverage varies monotonically with the miscoverage level \(\delta\), validating the finite-sample guarantees in practice.
3. Three rule families provide flexibility: The AND rule is most conservative (nearly all cases referred to X-ray), the OR rule is most lenient (roughly half of X-rays can be saved), and the Alpha-only rule falls in between.
4. Borderline cases are automatically referred to X-ray: Many borderline hips have α angles within a few degrees of the 60° threshold; the system does not force an ultrasound-only decision for these cases but instead refers them to X-ray, consistent with the behavior of experienced clinicians when uncertain.

Highlights & Insights

• Measurement-faithful rather than black-box: The entire reasoning chain is visible — landmark and point annotations → derived measurements → calibrated lower bounds → rule evaluation → policy outcome. There are no hidden logits or opaque multi-class outputs controlling safety-critical decisions, which is essential for clinical deployment.
• Label-efficient pipeline: Self-supervised pretraining leverages large volumes of unannotated data to learn representations; with the encoder frozen, only minimal annotations (30 subjects) are needed to train the head, perfectly suited to the reality of expensive annotation in medical imaging.
• Conformal prediction provides distribution-free guarantees: No assumptions about the data distribution are required; finite-sample coverage control is achieved under the exchangeability condition alone, making this more robust than conventional confidence intervals.
• Decision curves explicitly expose trade-offs: By parameterizing radiation cost \(\lambda\) and missed-finding penalty \(\mu\), modeling is decoupled from policy, placing radiation management decisions in the hands of the clinical team.
• Fail-safe design: Uncertain cases are automatically referred to X-ray rather than forced through an ultrasound-only decision; the one-sided conformal lower bound guarantees coverage control.

Limitations & Future Work

1. Small data scale: Annotated data comprise only 75 subjects, and the calibration set only 7 subjects (26 ultrasound images), which may result in wide conformal radii. Larger-scale, multi-center validation is needed.
2. Trainee annotators: Measurement annotations were produced by trainees rather than senior experts, potentially introducing annotation noise.
3. Age- and ossification-aware thresholds not implemented: Although an ossification nucleus landmark interface is provisioned, age-dependent differentiated thresholds are not actually implemented.
4. No prospective clinical study: Validation in actual clinical workflows — including radiation savings per 100 infants and changes in recall rates — is still needed.
5. Cross-modal proxy model not completed: The paper mentions the possibility of predicting X-ray AI or IHDI risk from ultrasound measurements as a proxy, but this extension is not implemented.
6. High MAE for femoral head coverage: The MAE for femoral head coverage is approximately 14 percentage points, leaving room for improvement.

Related Work & Insights

This paper is an excellent example of applying statistical decision theory (conformal prediction, decision curve analysis) to clinical imaging workflow optimization. Rather than pursuing peak accuracy on a single modality, it integrates cross-modal information into an actionable selective decision framework. This methodology generalizes to other clinical scenarios involving the question of "is this expensive test worth performing?" — for instance, performing a low-cost screening test first (blood panel, simple imaging) before deciding whether to proceed with an expensive or risky follow-up examination (contrast imaging, biopsy, etc.). The application of conformal prediction in medical AI is an emerging area, and this paper demonstrates a practical path for integrating it with clinical decision-making.

Rating

• Novelty: ⭐⭐⭐⭐ — First cross-modal selective imaging strategy grounded in conformal prediction
• Experimental Thoroughness: ⭐⭐⭐ — Small data scale; lacks external validation and prospective studies
• Value: ⭐⭐⭐⭐ — Highly actionable in clinical workflows; decisions are transparent and auditable
• Writing Quality: ⭐⭐⭐⭐⭐ — Methodology is articulated with exceptional clarity; clinical motivation and technical solution are tightly integrated

Related Papers

• [AAAI 2026] Learning with Preserving for Continual Multitask Learning
• [AAAI 2026] FaNe: Towards Fine-Grained Cross-Modal Contrast with False-Negative Reduction and Text-Conditioned Sparse Attention
• [AAAI 2026] Bidirectional Channel-selective Semantic Interaction for Semi-Supervised Medical Segmentation
• [CVPR 2026] CRFT: Consistent-Recurrent Feature Flow Transformer for Cross-Modal Image Registration
• [AAAI 2026] Measuring Stability Beyond Accuracy in Small Open-Source Medical Large Language Models for Pediatric Endocrinology