PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Md Shihabul Islam Shovo, Nishi Kanta Paul, Jamil Fayyad

arXiv:2610.29740v1cs.CV

TL;DR

Compact CXR localization is constrained by scarce box supervision and report-aware models that are too large or text-dependent for first-read inference. APED-CXR trains an image-only student using a frozen report-aware teacher to score the student’s current proposals; refreshed proposals improve held-out PadChest-GR grounding over fixed proposals while diagnostic AUC remains similar.

  • Problem

    Compact CXR models have limited box supervision, while report-aware multimodal models are expensive or text-dependent at inference.

    Source: §1 · p. 1

  • Method

    APED-CXR distills a frozen teacher’s report-informed scores on current student proposals into shared evidence maps for image-only diagnosis and grounding.

    Source: §1 · p. 1, §2.4 · p. 2, §2.3 · p. 2

  • Results

    APED-CXR reaches 0.338 micro-gF1, a paired gain of 0.036 over C2, while diagnostic AUC remains between 0.809 and 0.815 across branches.

    Source: §4.2 · p. 3

  • Takeaways & Limitations

    On held-out PadChest-GR, refreshed learner-derived proposals outperform fixed proposals and other proposal sources; lagged proposals recoup part of the gain, and the direction holds on NIH ChestX-ray14.

    Source: §5 · p. 4, §4.2 · p. 4

    The primary test set is from the training hospital, and NIH ChestX-ray14 is only a secondary cross-dataset check because it covers four classes without negatives and was used to train the teacher.

    Source: §4.3 · p. 4

Abstract

from arXiv · show

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.

1. INTRODUCTION AND RELATED WORK

Compact image-only CXR models need to localize findings despite limited box supervision, while report-aware multimodal models are too large or text-dependent for first-read use. APED-CXR distills report-informed spatial evidence into a student by having a frozen teacher verify student-proposed regions during training.

Motivation

  • Most training samples have labels and reports but lack boxes, despite evidence that image-level localization methods can underperform radiologists.PadChest-GR has radiologist boxes for 4,555 studies, compared with over 160,000 PadChest studies.Source: §1 · p. 1

Related Work

  • Report-aware CXR models can ground findings, but may require text at inference or billions of parameters, limiting their use as compact first-read tools.RegionCLIP instead uses a vision–language model to score region proposals as pseudo-labels for a separate model.Source: §1 · p. 1
  • Prior CXR distillation transfers labels, embeddings, or explanations rather than spatial evidence, leaving open which regions a teacher should evaluate.Fixed warm-up proposals may become mismatched as student evidence changes; learner-derived supervision's value for pathology localization remains unclear given query cost and imperfect box judgments.Source: §1 · p. 1

Contribution

  • APED-CXR uses a frozen multimodal teacher to verify current student proposals and distill report-informed support scores into evidence maps used for image-only diagnosis and grounding.Its evaluation uses a patient-disjoint held-out PadChest-GR set, five paired seeds, and patient-level bootstrap intervals, with a secondary NIH ChestX-ray14 evaluation.Source: §1 · pp. 1–2

2. METHOD

APED-CXR uses shared evidence maps for disease prediction and box generation, then distills a frozen teacher’s report-informed support for student-proposed regions. Its adaptive schedule refreshes those regions from the current student, unlike the static comparator.

2.1. Evidence-Map Student

  • An EfficientNet-B3 encoder and lightweight adapter produce a stride-8 evidence map, whose bounded class maps support disease probabilities through log-sum-exp pooling and box generation at inference.At inference, thresholded connected components become boxes after nonmaximum suppression; boxes are generated only when the class probability passes its threshold.Source: §2.1 · p. 2

2.2. Proposals and Privileged Verification

  • For each positive image–class pair, the student supplies up to three evidence-ranked components plus an exploration box, and the frozen MedGemma teacher scores each outlined region using the image, report, and class.The teacher is asked whether each region captures one visible occurrence of the finding; its answer-token likelihoods form the support score.Source: §2.2 · p. 2

2.3. Training Objective

  • Training combines classification BCE, box BCE-and-Dice supervision on manually boxed images, and soft-target BCE matching student proposal scores to teacher support scores.Uncertain and unmentioned report-derived labels are masked rather than treated as negative; filled boxes are coarse targets, not lesion masks.Source: §2.3 · p. 2

2.4. Adaptive and Static Proposal Schedules

  • The adaptive branch regenerates proposals from the current student every 1,200 updates, while the static branch keeps proposals from warm-up checkpoint S0 fixed; both share the same training budget and other settings.Student evidence is pooled again inside the fixed proposals at every update, and duplicate teacher queries are cached.Source: §2.4 · p. 2

3. EXPERIMENTAL SETUP

The study trains on a patient-stratified PadChest subset and evaluates primarily on held-out PadChest-GR, with NIH ChestX-ray14 as a secondary set. It compares adaptive proposals with controls and specifies box-level and diagnostic evaluation procedures.

3.1. Datasets

  • Training uses 40,000 patient-stratified PadChest images with positive target findings, while the 915-study PadChest-GR test partition is held out from training and tuning.Six findings are retained after label auditing; NIH ChestX-ray14 is secondary and is not independent of the teacher’s supervision because the teacher was trained on that dataset.Source: §3.1 · p. 2

3.2. Controls

  • Controls isolate query exposure, proposal coverage, scorer type, recency, teacher-generated targets, and annotation-based feedback against the adaptive current-proposal teacher branch.The annotation-feedback control is assessed only on grounding-train images.Source: §3.2 · p. 2, Figure 1 · p. 3

3.3. Evaluation Protocol

  • The primary metric is class-aware box-level micro-gF1 at IoU 0.5 on PadChest-GR, with uncertain, unmentioned, and non-localizable positives excluded.NIH is evaluated by class-conditional box recall for four findings; paired-difference confidence intervals use patient-level bootstrap samples conditional on trained seeds.Source: §3.3 · p. 3

3.4. Implementation Details

  • Training uses three stages of 40,000, 6,000, and 12,000 updates, with proposal refresh every 1,200 updates and teacher inference in bfloat16 without gradients.Experiments use PyTorch on one NVIDIA A100 GPU with 40 GB memory.Source: §3.4 · p. 3
  • Table 1 evaluates teacher scores on grounding-validation proposals using acceptability, AUROC, Top-1 acceptability, IoU rank correlation, all-bad support, and scoring time.Acceptability is same-class IoU of at least 0.5; generated boxes are mapped to proposals by overlap.Source: Table 1 · p. 3

4. RESULTS AND DISCUSSION

Adaptive teacher scoring of current student proposals yields stronger held-out grounding than fixed warm-up proposals, while diagnostic AUC remains stable across branches. The section also reports teacher-verification results, secondary evaluations, and limits on efficiency and external validation.

4.1. Teacher Verification

  • Combined image-and-report verification gives the best proposal AUROC (0.781) and top-1 hit rate (0.55), and wrong-location boxes score 0.24 lower on average.Source: §4.1 · p. 3

4.2. Adaptive versus Static Supervision

  • Full reaches 0.338 micro-gF1, a 0.036 gain over C2’s 0.302 (95% CI [0.022, 0.049]), while diagnostic AUC remains 0.809–0.815.Source: §4.2 · pp. 3–4, Table 2 · p. 4

4.3. Limitations and Efficiency

  • The student uses 11.1M parameters and takes 9.6 ms per image, including box extraction; adaptivity costs 34.4 teacher GPU-hours per seed, 5.8 times C2.Source: Table 3 · p. 4, §4.3 · p. 4

4.2. Adaptive versus Static Supervision

  • Random, dense, report-rule, and generated-box controls stay within 0.006 of C2, while current-proposal Full exceeds each by at least 0.034; two-refresh-old proposals recover only part of the gain.Source: §4.2 · p. 4
  • On NIH ChestX-ray14, adaptive supervision raises box recall from 0.423 to 0.447, while gains are largest for atelectasis, consolidation, and effusion.Source: §4.2 · p. 4

4.3. Limitations and Efficiency

  • The primary test set comes from the training hospital, and NIH is only a secondary cross-dataset check because it covers four classes without negatives and was used to train the teacher.Source: §4.3 · p. 4

4.2. Adaptive versus Static Supervision

  • Grounding-val C2 saturates after roughly two refresh intervals, whereas Full and lagged C7 continue improving, with C7 about two refreshes behind.Source: Figure 2 · p. 4, §4.2 · p. 4

5. CONCLUSION

APED-CXR transfers report-informed spatial evidence through teacher verification of the student’s own proposals. Refreshed proposals improved grounding on held-out PadChest-GR relative to fixed and other proposal sources, with the direction also holding on NIH ChestX-ray14.

5. CONCLUSION

  • On held-out PadChest-GR, refreshed learner-derived proposals outperformed fixed proposals and other proposal sources on box-level grounding.The lagged proposals recouped part of the gain.Source: §5 · p. 4
  • The direction of the grounding result also held on NIH ChestX-ray14.Source: §5 · p. 4
  • Future work will reduce teacher cost through selective refreshing and use larger multi-institution test sets.Source: §5 · p. 4

6. COMPLIANCE WITH ETHICAL STANDARDS

The study used de-identified data from PadChest, PadChest-GR, and NIH ChestX-ray14 under their research use agreements. Ethical approval was not required for this retrospective study, and no PadChest images or reports are reproduced.

6. COMPLIANCE WITH ETHICAL STANDARDS

  • PadChest and PadChest-GR came from BIMCV, while NIH ChestX-ray14 was downloaded from the NIH site.Source: §6 · p. 5
  • Ethical approval was not required for this retrospective study of de-identified data under the source datasets’ data use agreements.Source: §6 · p. 5
  • The paper reproduces a public NIH ChestX-ray14 radiograph in Fig. 1 and no PadChest images or reports.Source: §6 · p. 5

Related papers

Loading 2610.29740v1…