JDCNet: Confidence-Gated Privileged-Modality Distillation for Cost-Preserving X-ray Inference

Bo Ma; Jinsong Wu; Weiqi Yan; Hongjiang Wei; Kun Liu

cs.CVMar 31, 2026

JDCNet: Confidence-Gated Privileged-Modality Distillation for Cost-Preserving X-ray Inference

Authors: Bo Ma, Jinsong Wu, Weiqi Yan, Hongjiang Wei, Kun Liu

Organizations: J. Wu is with the Guilin University of Electronic Technology, Guilin, China. · H. Wei is with Hikvision Technology Co., Ltd, Hangzhou, China. · K. Liu is with the Hebei University of Technology, Tianjin 300401, China.

Abstract

We study a systems-level visual inference problem: using an expensive privileged modality during training while preserving a fixed-cost, single-modality deployment path. We present JDCNet, a confidence-gated CT-to-X-ray distillation framework in which the CT teacher supplies an auxiliary hard or temperature-scaled target only on training samples whose teacher confidence exceeds a threshold; at deployment the student takes X-ray input alone and matches the parameter, MAC, and latency profile of the supervised X-ray baseline. On a 510-patient same-patient paired BIMCV cohort with patient-level 5-fold cross-validation, two JDCNet configurations clear a fixed transfer gate against the supervised ResNet-18 baseline: 3-slice soft-KL supervision yields $Δ\mathrm{BA}{=}{+}0.035$ ( $95\%$ CI $[{+}0.011,{+}0.057]$ ) and mid-slice hard supervision yields $+0.033$ ( $[{+}0.007,{+}0.058]$ ). Under the same splits and gate, logit distillation, gated logit distillation, contrastive alignment, attention transfer, feature hints, BiomedCLIP fine-tuning, and a module-augmented variant do not pass. Confidence-gated auxiliary targets are therefore a more transferable channel than uniformly softened CT logits; the evidence is bounded to one paired cohort, so external paired-cohort replication is required before any deployment claim.

Figures & tables

Fig. 1: JDCNet mechanism diagram. Top: at training time, the CT teacher branch produces softmax probabilities $p^{ct}$ , the confidence mask $\mathcal{M}$ retains only the samples whose top-class probability exceeds $\tau_{\mathrm{gate}}$ , and either the argmax (hard variant) or the temperature-scaled distribution (soft-KL variant) supervises the X-ray student in addition to the always-on weighted cross-entropy on the ground-truth label. Gradients flow only through the student parameters (red dashed). Bottom: at deployment, the CT input, CT encoder, and confidence-gated auxiliary loss are dropped; the deployed X-ray classifier is structurally identical to a standalone supervised ResNet-18, so JDCNet adds zero inference-time cost.

Fig. 2: Evidence map for the comparator distillation families evaluated under the early ten-resample same-case protocol on the smaller paired cohort. Bars show mean balanced accuracy with one standard deviation across patient-level Monte Carlo resamples. The dashed line marks the X-ray-only student mean. Plain cross-modal logit KD and cross-modal attention transfer are the strongest comparator means, but none of these mechanisms clears the fixed 510-patient transfer gate (Section IV ); the validated JDCNet method introduced in Section III-C is evaluated separately on the 510-patient cohort. Generic KD baselines (CRD, DKD, DIST, modality hallucination) cluster below the supervised baseline, and the full DPE/MHRA/DFPN module stack does not improve upon stripped-down logit transfer.

Question	Method / view	$\Delta\mathrm{BA}$ [95% CI]	Cells	Interpretation
Teacher upper bound	CT mid-slice	$+0.045$ [ $+0.019,+0.069$ ]	11/15	CT carries additional patient-level signal.
Teacher upper bound	CT 3-slice	$+0.051$ [ $+0.025,+0.080$ ]	12/15	Strongest teacher upper-bound row.
Validated transfer	JDCNet, 3-slice soft-KL	$+0.0345$ [ $+0.0112,+0.0571$ ]	10/15	Passes the fixed transfer gate.
Validated transfer	JDCNet, mid hard	$+0.0329$ [ $+0.0074,+0.0584$ ]	10/15	Independent positive JDCNet setting.
Comparator	Gated logit KD, mid	$-0.022$ [ $-0.055,+0.011$ ]	–	Confidence weighting alone is insufficient.
Comparator	Gated logit KD, DRR	$-0.064$ [ $-0.095,-0.034$ ]	–	Geometric projection does not rescue logit transfer.

TABLE I: Main 510-patient decision summary. All deltas are paired balanced-accuracy differences against the matched supervised X-ray baseline unless marked as teacher upper bound. The fixed gate is mean

\Delta\mathrm{BA}\geq+0.03

with bootstrap-CI lower bound

>0

Cohort / source	Support	Pairing	Role in paper	Main caveat
Primary BIMCV paired cohort	510 patients ( $113^{+}/397^{-}$ )	Same-patient CT–X-ray	Headline validation for JDCNet and matched comparators	Single public paired cohort; binary COVID/non-COVID task.
Historical Cohen paired subset	19 patients / 26 X-rays ( $22^{+}/4^{-}$ images)	Same-patient CT–X-ray	Feasibility and instability diagnosis only	Too small for headline ranking or threshold-sensitive metrics.
Cohen all-X-ray / all-CT references	424 X-ray patients; 25 CT patients	Not same-case comparable	Single-modality data-availability references	Not a deployment-realistic CT-to-X-ray comparison.
Exploratory stress-test cohorts	226/228/512-patient BIMCV variants, MIDRC pilots, cross-source controls	Varies	Appendix stress tests and failure-mode diagnostics	Not used to establish the main JDCNet claim.

TABLE II: Dataset roles after revision. Only the 510-patient same-patient paired BIMCV cohort supports the headline JDCNet claim; other cohorts are historical, reference-only, or exploratory stress tests.

Evidence layer	Main observation	Role
Small paired fixed split	4 held-out paired patients, severe class imbalance	Feasibility debug only
226-patient calibration scan	Best gated-logit cell positive mean, CI crosses zero	Instability diagnosis
510-patient logit-KD extension	Gated logit KD negative or neutral across teacher views	Mechanism audit
Cross-source controls	Source shift can erase specificity	External-validity warning
Foundation-model control	BiomedCLIP tied with ResNet-18 supervised X-ray	Capacity control

TABLE III: Historical and stress-test evidence retained for interpretation. These rows diagnose instability, source shift, and capacity controls; they do not establish the headline JDCNet claim.

Row	Params/MACs	Mem.	CPU (bs=1)	GPU (bs=1)	Edge (bs=1)
Supervised X-ray	11.18 M / 1.82 G	38 MB	21.2 (0.6) ms	1.85 (0.04) ms	18.7 (0.5) ms
JDCNet student	11.18 M / 1.82 G	38 MB	21.2 (0.6) ms	1.85 (0.04) ms	18.7 (0.5) ms
Plain/gated logit-KD	11.18 M / 1.82 G	38 MB	21.2 (0.6) ms	1.85 (0.04) ms	18.7 (0.5) ms
Module-augmented pilot	0.57 M / 3.05 G	71 MB	45.1 (1.4) ms	3.92 (0.09) ms	41.4 (1.2) ms

TABLE IV: Deployment cost at

224{\times}224

FP32 input. Hardware: CPU = Intel Xeon Silver 4210 (single thread, BLAS pinned, no SIMD overrides); GPU = NVIDIA RTX 3090 24 GB; Edge = NVIDIA Jetson Orin Nano 8 GB, 15 W power mode. Software stack: Ubuntu 22.04, PyTorch 2.1.0 + CUDA 12.1 (server) and JetPack 6.0 + TensorRT 8.6 (edge). Each cell is median (IQR) of 1000 timed forward passes after 50 warm-ups at batch size 1, including the

224{\times}224

resize and per-channel normalization (no test-time augmentation). Peak GPU activation memory was sampled with torch.cuda.max_memory_allocated . Rows 1–3 share the same ResNet-18 X-ray inference graph, so JDCNet adds zero inference-time cost relative to supervised X-ray.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Teacher	$\tau$	$\lambda$	Target	Pos./15	$\Delta\mathrm{BA}$ [95% CI]	Gate
3-slice	0.70	1.0	soft-KL	10	$+0.0345$ [ $+0.0112$ , $+0.0571$ ]	Pass
Mid-slice	0.80	1.5	hard	10	$+0.0329$ [ $+0.0074$ , $+0.0584$ ]	Pass
Mid-slice	0.70	1.0	hard	10	$+0.0298$ [ $-0.0002$ , $+0.0597$ ]	–
Mid-slice	0.70	1.5	hard	10	$+0.0296$ [ $+0.0026$ , $+0.0577$ ]	–
Mid-slice	0.70	1.0	soft-KL	10	$+0.0264$ [ $+0.0017$ , $+0.0547$ ]	–
3-slice	0.80	1.0	soft-KL	10	$+0.0258$ [ $-0.0029$ , $+0.0546$ ]	–

Appendix

TABLE I: Full JDCNet sweep on the 510-patient BIMCV paired cohort. Each row uses the same ResNet-18 X-ray student, patient-level 5-fold cross-validation, and seeds 42–44. The fixed gate is mean

\Delta\mathrm{BA}\geq+0.03

and bootstrap-CI lower bound

>0

Method	BA	AUC	F1	Sens.	Spec.
Supervised X-ray	$.604{\pm}.038$	$.661{\pm}.041$	$.566{\pm}.040$	$.547{\pm}.072$	$.661{\pm}.054$
JDCNet 3-slice soft-KL	$.639{\pm}.034$	$.701{\pm}.038$	$.598{\pm}.036$	$.598{\pm}.066$	$.679{\pm}.049$
JDCNet mid hard	$.637{\pm}.037$	$.696{\pm}.040$	$.595{\pm}.038$	$.591{\pm}.069$	$.682{\pm}.051$
Gated logit KD (mid)	$.582{\pm}.045$	$.642{\pm}.048$	$.547{\pm}.045$	$.529{\pm}.078$	$.634{\pm}.061$
Gated logit KD (DRR)	$.540{\pm}.049$	$.598{\pm}.052$	$.508{\pm}.048$	$.512{\pm}.082$	$.568{\pm}.066$
Contrastive align. (best)	$.612{\pm}.041$	$.668{\pm}.044$	$.573{\pm}.042$	$.555{\pm}.073$	$.669{\pm}.056$

Appendix

TABLE II: Absolute performance on the 510-patient BIMCV paired cohort, mean

\pm

SD across 15 fold/seed cells. Sens. / Spec. = sensitivity (COVID-positive) / specificity (non-COVID).

Row	$\Delta\mathrm{BA}$	Fold/seed CI	Patient-level CI
CT teacher, mid (u.b.)	$+0.045$	$[+0.019,+0.069]$	$[+0.012,+0.078]$
CT teacher, 3-slice (u.b.)	$+0.051$	$[+0.025,+0.080]$	$[+0.017,+0.087]$
JDCNet, 3-slice soft-KL	$+0.0345$	$[+0.0112,+0.0571]$	$[+0.0078,+0.0613]$
JDCNet, mid hard	$+0.0329$	$[+0.0074,+0.0584]$	$[+0.0041,+0.0625]$
Gated logit KD, mid	$-0.022$	$[-0.055,+0.011]$	$[-0.064,+0.019]$
Gated logit KD, DRR	$-0.064$	$[-0.095,-0.034]$	$[-0.108,-0.022]$

Appendix

TABLE III: Fold/seed and patient-level paired bootstrap 95% CIs for the headline rows and the strongest failed comparators. Patient-level resampling is the conservative inference layer; the gate decision is unchanged.

Teacher / $\tau$	Cov.	Cov. +	Cov. -	$\Delta$ Acc.	$\Delta\mathrm{BA}$
Mid, 0.60	0.82	0.78	0.84	$+0.08$	$+0.018$
Mid, 0.70	0.73	0.69	0.75	$+0.10$	$+0.026$
Mid, 0.80	0.63	0.58	0.65	$+0.13$	$+0.033$
Mid, 0.85	0.55	0.49	0.58	$+0.14$	$+0.027$
3-slice, 0.60	0.84	0.80	0.85	$+0.09$	$+0.020$
3-slice, 0.70	0.74	0.69	0.76	$+0.11$	$+0.035$

Appendix

TABLE IV: Gate-coverage and teacher-reliability diagnostics across the four CT teacher views. Cov. = fraction of training samples retained by the confidence mask; Cov. + /Cov. - = per-class retention;

\Delta

Acc. = retained-minus-rejected teacher accuracy.

Family	Representative result	Gate
CT teacher u.b.	Mid $+0.045$ [ $+0.019,+0.069$ ]; 3-slice $+0.051$ [ $+0.025,+0.080$ ]	Pass
JDCNet	3-slice soft-KL $+0.0345$ [ $+0.0112,+0.0571$ ]; mid hard $+0.0329$ [ $+0.0074,+0.0584$ ]	Pass (2/16)
Gated logit KD	Mid $-0.022$ [ $-0.055,+0.011$ ]; DRR $-0.064$ [ $-0.095,-0.034$ ]	Fail
Contrastive align.	Best cell $+0.008$ [ $-0.020,+0.037$ ]	Fail
BiomedCLIP fine-tune	$+0.002$ [ $-0.048,+0.050$ ]	Fail
Historical 226 ref.	Near-pass $+0.034$ , CI crosses zero	Hist.

Appendix

TABLE V: Comparator interpretation under the same paired decision rule.

Teacher	$\tau$	BA mean [95% CI]	$\Delta\mathrm{BA}$ [95% CI]	Gate
3-slice stack	0.20	0.633 [0.610, 0.656]	$+0.008$ [ $-0.020,+0.037$ ]	Fail
3-slice stack	0.07	0.627 [0.594, 0.657]	$+0.003$ [ $-0.023,+0.030$ ]	Fail
Mid-slice	0.20	0.620 [0.594, 0.643]	$-0.005$ [ $-0.031,+0.020$ ]	Fail
Mid-slice	0.07	0.616 [0.590, 0.642]	$-0.008$ [ $-0.038,+0.027$ ]	Fail

Appendix

TABLE VI: Cross-modal contrastive alignment (CLIP-style InfoNCE pretrain + supervised fine-tune) on the 510-patient BIMCV paired cohort. Each row is

n=15

paired fold/seed cells.

\Delta\mathrm{BA}

is the paired difference against the supervised X-ray baseline.

Method	Sensitivity	Specificity	ROC-AUC
Supervised X-ray	$1.00{\pm}.00$	$0.32{\pm}.34$	$0.32{\pm}.34$
Plain cross-modal logit KD	$1.00{\pm}.00$	$0.00{\pm}.00$	$0.00{\pm}.00$
Attention transfer	$1.00{\pm}.00$	$0.10{\pm}.25$	$0.10{\pm}.25$
Feature hints	$1.00{\pm}.00$	$0.00{\pm}.00$	$0.00{\pm}.00$
Module-augmented pilot	$1.00{\pm}.00$	$0.27{\pm}.38$	$0.27{\pm}.38$

Appendix

TABLE VII: Cross-source non-COVID control. Mean

\pm

std over 8 patient-level resamples. ROC-AUC equals specificity because probability outputs become near-binary under distribution shift.

Teacher	Method	$\Delta\mathrm{BA}$ [95% CI]	Gate
Mid-slice	Teacher	$+0.012$ [ $-0.024,+0.046$ ]	—
	Plain KD	$-0.022$ [ $-0.062,+0.021$ ]	Fail
	Gated KD	$-0.014$ [ $-0.053,+0.023$ ]	Fail
3-slice stack	Teacher	$+0.019$ [ $-0.013,+0.058$ ]	—
	Plain KD	$-0.030$ [ $-0.083,+0.016$ ]	Fail
	Gated KD	$-0.045$ [ $-0.092,+0.000$ ]	Fail

Appendix

TABLE VIII: CT teacher representation comparison on the BIMCV-only balanced 226-patient 5-fold CV (

n{=}15

cells per row).

\Delta\mathrm{BA}

is the paired difference against the supervised baseline trained under the same representation-specific split.

Cross-Modal Knowledge Distillation

Cross-modal knowledge distillation transfers useful representations or predictions between modalities, often enabling cheaper single-modality inference.

See all 122 papers

APED-CXR: Adaptive Privileged Evidence Distillation for Chest X-Ray Grounding with Image-Only Inference
Oct 21, 2026Md Shihabul Islam Shovo, Nishi Kanta Paul, Jamil FayyadMedical Image Grounding Adaptive Knowledge Distillation
NIMISHES Lab, Dhaka, Bangladesh · American University of Sharjah, Sharjah, UAE
Cross-Modal Guidance for Fast Diffusion-Based Computed Tomography
Mar 1, 2026Timofey Efimov, Singanallur Venkatakrishnan, Maliha Hossain +2Computed Tomography Diffusion Model Guidance
Carnegie Mellon University, Pittsburgh, USA · Oak Ridge National Laboratory, Oak Ridge, USA · Purdue University, West Lafayette, USA
JANUS: Anatomy-Conditioned Gating for Robust CT Triage Under Distribution Shift
May 13, 2026Lavsen Dahal, Yubraj Bhandari, Geoffrey Rubin +1Distribution Shift Robustness Healthcare
Center for Virtual Imaging Trials, RAI Labs, Department of Radiology, Duke University, Durham NC 27708, USA · Electrical and Computer Engineering, Pratt School of Engineering, Duke University, Durham, NC 27708, USA · Department of Mathematics, Trinity College of Arts & Sciences, Duke University,2026 Durham, NC 27708, USA +1

Oct 21, 2026J/K move · Enter open · S save

APED-CXR: Adaptive Privileged Evidence Distillation for Chest X-Ray Grounding with Image-Only Inference

Md Shihabul Islam Shovo, Nishi Kanta Paul, Jamil Fayyad

NIMISHES Lab, Dhaka, Bangladesh · American University of Sharjah, Sharjah, UAE

Medical Image Grounding Adaptive Knowledge Distillation Radiology VLMs Chest X-Ray Classification

Open paper AskPDF arXiv

Compact chest X-ray (CXR) models that localize findings suit deployment, but box annotations are limited and report-aware multimodal models are expensive at inference. We introduce APED-CXR, which distills report-informed spatial evidence from a frozen multimodal teacher into a student with 11.1M parameters that predicts diseases and boxes from shared evidence maps and requires the image alone at test time. During training, the teacher uses the radiograph, report, and queried class to evaluate regions proposed from the student's current evidence maps, and its support scores serve as training targets for the student's pooled evidence scores within these regions. Trained on PadChest and evaluated on held-out PadChest-GR, refreshing the student's proposals for teacher evaluation improves box-level localization F1 from 0.302 with fixed warm-up proposals to 0.338 (IoU >= 0.5; gain 0.036, 95% CI [0.022, 0.049]), while maintaining diagnostic AUC. This gain depends on the teacher checking current proposals: with equally many random boxes, a fixed dense box set, teacher-generated boxes, or report rules instead of the teacher, F1 stays near the fixed-proposal level, and proposals two refreshes old recover only part of the gain. On NIH ChestX-ray14, a secondary test set, box recall also rises from 0.423 to 0.447.

Abstract

Cross-Modal Knowledge Distillation

Explore similar work

APED-CXR: Adaptive Privileged Evidence Distillation for Chest X-Ray Grounding with Image-Only Inference

Cross-Modal Guidance for Fast Diffusion-Based Computed Tomography

JANUS: Anatomy-Conditioned Gating for Robust CT Triage Under Distribution Shift