Pith.

Machine Learning · arXiv:2609.26384 v1 · submitted 22 Sep 2026 · 13 references

Learning to Defer with Guidance on Real World Medical Data

Emma Sun, Joshua Strong, Alison Noble

The paper says

“Medical image interpretation is high-volume and time-consuming, and while AI interpretation can reduce workload, fully autonomous deployment carries potential safety concerns and low specificity may in practice lead to increased clinician workload.”

more of the abstract

Learning to Defer (L2D) addresses this by selectively routing cases between autonomous prediction and human experts by learning from input features and AI model and human performance. While theoretical guarantees have been proven for L2D, its performance has not been validated on real-world medical datasets with human reader annotations. We evaluate the predictor-rejector formulation of two-stage L2D, where the AI predictor model is fixed and separate from the trainable routing or rejector model, on Collab-CXR, a multilabel chest X-ray dataset with multiple human annotations per case. This is the first work to look at L2D in the context of real-world medical imaging data with human annotations. We further introduce a new setup, L2D with Guidance, where the decision space is extended to three choices: predict autonomously, defer to a human expert, or defer to a human expert and provide AI guidance. We compare multiple rejector architectures and loss functions, and different input feature availabilities. This is reproduced on two larger datasets, VinDr-CXR and CheXpert. Our results show that two-stage L2D with Guidance outperforms classic two-stage learning to defer, as well as human-alone, AI-alone and AI-guided human baselines. Notably, this performance is achieved with simpler loss functions compared to formally defined L2D surrogate loss functions in current literature.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 24 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. First real human-annotated medical test of two-stage L2D with a genuinely new three-way defer/guidance action, but the headline outperformance rests on a pooled meta-expert baseline that needs a per-reader check before it transfers to any real clinician. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 13 references say.

The paper says, at Sec. 3.1 (b44–b45), Table 2

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The human-alone (0.919) and AI-guided-human (0.923) baselines on Collab-CXR are accuracies of a pooled meta-expert formed from up to 10 inconsistently-assigned radiologists. No single real reader necessarily matches this pooled prediction, and an ensemble can exceed every member's accuracy, so the headline comparison (0.975 vs 0.923, b9) is against a fictional expert that no deployable clinician would match.

more of objection 1

The paper acknowledges the pooling necessity (b44, b45) but argues only that pooling preserves the difference between the two human conditions, not that the pooled accuracy level is representative. Please report the distribution of individual radiologist accuracies (overall and per pathology where the annotation scheme permits) and verify that L2D-G also outperforms the best available single reader, or explicitly restrict the claim to 'outperforms the pooled expert' and discuss the deployment implications of routing cases to a reader whose correctness pattern differs from the pooled target the rejector was trained on.

The paper says, at Sec. 3.1 (b54), Table 2

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The CheXpert and VinDr-CXR reproductions simulate the AI and AI-guided human experts using pathology-specific flip probabilities that are tuned so that the baselines land within ±0.9% of each other and relative performance matches Collab-CXR. These simulation parameters are free choices that directly determine the gap the rejector is meant to exploit, so the reported margins (0.988 vs 0.901 on CheXpert; 0.969 vs 0.939 on VinDr-CXR) may be inflated relative to what would be observed under other plausible expert models.

more of objection 2

Please add a sensitivity analysis varying the flip probabilities within reasonable ranges and report how the L2D-G advantage changes; without this, the reproduction leg remains a consistency check with tuned experts rather than an independent validation.

The paper says, at Sec. 4, statistical reporting

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The paper reports paired t-tests for the MLP-vs-formal-L2D comparison (p=0.022) and for L2D-G vs classic L2D (p<0.0001) on Collab-CXR, which is good, but no significance values are given for the headline L2D-G vs baselines comparisons on any dataset, nor for the CheXpert/VinDr-CXR reproductions (b84). Given that the main contribution is the claim of outperformance over human and AI baselines, please provide explicit statistical comparisons (paired tests over the 20 seeds) for these differences as well.

5 minor objections

The paper says, at Sec. 3.1, Sec. 4

Pith says: minor objection 1 of 5

The accuracy metric is referred to as 'per-label accuracy' in the conclusion (b86) but is not formally defined. Please state explicitly whether accuracy is averaged over the 14 pathologies per image, or over image-level multilabel predictions, and whether it is computed per pathology-label or per image.

The paper says, at Fig. 3, Fig. 4

Pith says: minor objection 2 of 5

The y-axis is truncated in both figures, which can exaggerate visual differences. Please show the full axis range or add an explicit annotation of the axis limits; the caption already notes truncation, but showing the full range would make the baselines-rejector gap easier to assess.

This objection may describe the copy Pith gave the reviewer, not the paper.

show 3 more

The paper says, at Table 1

Pith says: minor objection 3 of 5

The '—' entries for tree-based models under 'Image features' should be explained in the text or caption; it is currently not stated whether this means image features were not evaluated for Random Forest/XGBoost or were evaluated and omitted from the table.

The paper says, at Sec. 3.1 (b46)

Pith says: minor objection 4 of 5

The statement that pooling is 'more realistically-grounded in human performance characteristics compared to synthetic annotations' is arguable, since the pooled meta-expert is also an artificial construction. Consider softening or supporting this with a per-reader analysis.

The paper says, at Sec. 3.1

Pith says: minor objection 5 of 5

The AI predictor is described as 'the AI tool from the dataset'; please identify it explicitly as the CheXbert labeller output, which is mentioned earlier but not consistently in the method description.

What would settle it

For each of the up to 10 Collab-CXR radiologists, compute per-pathology accuracy on the 14 used pathologies against ground truth, on the subset they read, both unguided and AI-guided. Then compare the pooled baselines (0.919 human-alone, 0.923 AI-guided) to the distribution (min, median, max) of individual accuracies. If either pooled baseline exceeds the maximum individual accuracy, or sits above the 75th percentile, the meta-expert is not representative of any single clinician and the outperformance claim must be reworded. Additionally, evaluate the best rejector configuration with test-time deferral restricted to a randomly chosen individual reader (instead of the pooled expert) to see whether the reported 0.975 accuracy decays materially. This directly tests whether the claimed advantage survives contact with a real radiologist.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 24 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 25 s · 9,031 tokens read, 3,496 written
  3. Editor wrote the verdict above 16 s · 10,100 tokens read, 1,623 written
  4. Referee major revision: 3 major objections, 5 minor 30 s · 9,381 tokens read, 4,290 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Read it. The genuinely new thing is empirical: two-stage L2D evaluated on real human-annotated chest X-rays, plus the three-way action space that lets the router defer to an AI-guided human instead of only predicting or deferring. On Collab-CXR, a simple MLP rejector reaches 0.975 vs a best baseline of 0.923 while deferring only 38% of cases. That is a real result, though a modest one. The paper also deserves credit for reporting 20 seeds, paired t-tests, and a sensible ablation of tabular versus image features, and for being honest that the dataset is small and the study design forced pooling.

The soft spots are in proportion to how much the central claim leans on them. The stress-test concern lands. The human expert is a pooled meta-reader of up to 10 inconsistently-assigned radiologists (b44, b45). Pooling can produce an accuracy higher than any individual member, and the paper reports only the pooled baselines, with no per-reader range. So the claim that L2D-G outperforms "human alone" is not yet a claim about any actual radiologist a deployable system would face. The authors acknowledge the pooling but argue only that it preserves the gap between conditions; that does not establish that the pooled accuracy level is representative. This is load-bearing, not a footnote.

The reproduction leg is also weaker than it looks. CheXpert and VinDr-CXR have real individual annotations, but the AI and AI-guided human outputs are simulated with flip probabilities calibrated so baselines land near each other in the Collab-CXR style. That makes the second and third datasets a consistency check of the method under favorable simulation, not an independent confirmation of the headline. And with only 324 images and 14 of 106 pathologies, the Collab-CXR finding is narrow even on its own terms. One smaller oddity: classic L2D fails to beat baselines on Collab-CXR but succeeds on the simulated datasets, and the paper's overfitting explanation is plausible but not verified.

Who is this for? People working on clinical workload routing or human-AI collaboration who want a first real-data data point for L2D. It is useful, clearly written, and the negative result for the formal surrogate loss is worth knowing. It is not yet an established outperformance claim.

Recommendation: send it to peer review. It deserves a serious referee. But the revision should require per-reader accuracy breakdowns on Collab-CXR, code and data release, and reframing of the simulated datasets as a sanity check rather than reproduction.

The strongest objection

The paper's central claim (b9) that L2D-G outperforms human-alone (0.919) and AI-guided human (0.923) rests on baselines built by pooling up to 10 inconsistently-assigned radiologists into a single meta-expert (b44, b45). The authors acknowledge the pooling but argue only that it preserves the difference between the two human conditions, not that the pooled accuracy level is representative of any actual reader. Pooling can produce an ensemble whose accuracy exceeds every member; if the pooled baseline is higher than the best individual radiologist, then the 0.975 vs 0.923 comparison is against a fictional entity that no deployable clinician matches. The rejector is trained to route to this meta-expert, and at deployment a real system would face one human reader, not the pool. If the pooled accuracy overstates individual human performance, the headline outperformance does not transfer. The paper even lists the pooled accuracies as baselines without reporting the range across individual readers, so the 'human-alone' baseline is load-bearing but unverified.

The reader's weakest_assumption identifies the same load-bearing concern I do: the pooled meta-expert is not a faithful proxy for any single radiologist, and the baselines of 0.919/0.923 may not reflect actual human performance. The paper explicitly acknowledges the pooling (b44, b45) but only defends it as a way to enable training and to preserve the difference between unguided and AI-guided readers; it does not verify that the pooled accuracy level is achievable by any individual. Because the rejector is trained to route to this pooled entity, the central outperformance claim depends on this unverified assumption. The concrete per-reader accuracy comparison would settle whether the 0.975 result is meaningful against a real clinician. This is not an internal inconsistency in the method, but an external validity gap in the baseline; it does not invalidate the approach, but it conditions the generality of the claim. The CheXpert/VinDr simulations are calibrated to Collab-CXR relative performance, so they do not independently resolve this concern. No other concern is more central: the small dataset size is acknowledged (b92), and the rejector's best accuracy does not rely on expert-output leakage since the 'Pathology' and 'Pathology+AI output' configurations also achieve 0.975 (Table 1). Therefore I agree with the reader and recommend no change to the CONDITIONAL verdict; the condition should explicitly require reporting the per-reader accuracy distribution.

The referee's report

If the results hold, this is a meaningful first step for L2D on real human-annotated medical data, and the L2D-G formulation with an explicit AI-guidance deferral option is a genuinely new extension. The strengths are the 20-seed experimental protocol with paired t-tests, the clear ablation over input features, and the instructive finding that a simple tabular MLP with a plain BCE loss outperforms the formal surrogate-loss rejector. The main reservation is that the Collab-CXR human baselines are pooled meta-expert accuracies, not the accuracy of any single radiologist, which limits the transferability of the central outperformance claim to a deployable system with one human reader.

Pith's reading of the paper (a machine's summary, not the authors')

The paper asks whether Learning to Defer (L2D), routing each case to an AI, a human expert, or in the new setup a human expert receiving AI guidance, can beat any single reader on real medical imaging data with genuine human annotations, a setting the L2D literature has previously only simulated. On Collab-CXR, a small multi-label chest X-ray dataset annotated by up to ten radiologists, the best router (a simple multilayer perceptron over tabular features) reaches 97.5% per-label accuracy while deferring 38.2% of cases, against a best baseline of 92.3% for the AI-guided human. The direction of the finding reproduces on two larger datasets, CheXpert and VinDr-CXR. If correct, the result shows that a fixed AI predictor plus a trainable router can outperform every individual reader in a real human-annotated setting, and that adding an "AI guidance" option to the deferral decision, rather than adopting a fancier surrogate loss, is what carries the gain.

What carries the argument. The load-bearing object is the rejector, a separate trainable router layered over a frozen AI predictor, which is the two-stage predictor-rejector formulation of L2D. The paper's contribution is a third action in the decision space: rather than only choosing between the AI's own output and an unguided human expert, the rejector may also defer to a human expert who is shown AI guidance. The rejector learns, per input, which of the three actions minimises expected cost, thereby exploiting cases where the AI, a plain human, or an AI-assisted human each have relative strengths. Four rejector architectures are compared (Random Forest, XGBoost, an MLP with binary cross-entropy, and an MLP with the formal surrogate loss from the two-stage L2D literature), alongside three input feature regimes (tabular only, tabular plus naive image embeddings, and tabular plus pathology-specific embeddings).

Core claim. On the paper's own terms, the discovery is that extending two-stage L2D's decision space from {predict, defer} to {predict, defer, defer with AI guidance} produces a system whose Collab-CXR accuracy exceeds every baseline it is compared against: AI alone (0.920), human alone (0.919), and AI-guided human (0.923). The best configuration is a simple MLP rejector trained with plain binary cross-entropy on tabular inputs (pathology labels plus AI output), reaching 0.975 ± 0.005 while deferring 38.2% of cases. Two findings follow on Collab-CXR: the formally defined L2D surrogate loss is outperformed by the simpler loss (paired t-test, p = 0.022), and classic two-stage L2D without the guidance option falls below the baselines, which the authors attribute to overfitting on the moderate-size dataset. The direction of the results carries over to CheXpert and VinDr-CXR, where classic L2D does beat baselines but L2D-G still wins.

The premise it rests on. The load-bearing premise is that pooling up to ten inconsistently-assigned radiologists into a single meta-expert preserves a faithful signal of expert performance for the router to learn from; the paper routes to a fictional average reader, and if no real radiologist matches that pooled entity, the reported outperformance over 'human-alone' would not transfer to an actual deployed clinician, a limitation the authors acknowledge in describing the pooling necessity.

If the paper is right (Pith's inferences, not the authors')

L2D with Guidance outperforms classic two-stage L2D, human-alone, AI-alone, and AI-guided human baselines on Collab-CXR, with the direction reproduced on CheXpert and VinDr-CXR.

A simple MLP rejector trained with binary cross-entropy on tabular inputs is sufficient, and it outperforms the formally defined L2D surrogate loss (paired t-test, p = 0.022).

Adding image features, whether naive embeddings or pathology-specific ones, does not improve the best tabular-only rejector on Collab-CXR.

Classic two-stage L2D without the guidance option fails to beat the baselines on the real-world Collab-CXR data, though it does beat them on the two larger simulated datasets.

The system reduces workload: on Collab-CXR only 38.2% of cases are deferred to human review while overall accuracy rises above every individual reader.

As my own extension: the three-way routing principle should generalise beyond chest X-ray to any workflow with a fixed, costly-to-retrain AI and heterogeneous clinician availability, such as out-of-hours reading where specialist coverage varies; the gain is likely largest where the AI and the available human have complementary error patterns.

One testable extension I would pursue: because the router targets a pooled meta-expert rather than any single real reader, a decisive check is to train one rejector per actual radiologist and see whether any real reader individually reproduces the 97.5%-versus-92.3% margin.

On CheXpert and VinDr-CXR the AI-guided human is itself simulated, so the guidance benefit on those datasets rests on the realism of that simulation; gathering genuine AI-assisted annotations there would test whether the benefit survives real assisted reading.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Medical image interpretation is high-volume and time-consuming, and while AI interpretation can reduce workload, fully autonomous deployment carries potential safety concerns and low specificity may in practice lead to increased clinician workload. Learning to Defer (L2D) addresses this by selectively routing cases between autonomous prediction and human experts by learning from input features and AI model and human performance. While theoretical guarantees have been proven for L2D, its performance has not been validated on real-world medical datasets with human reader annotations. We evaluate the predictor-rejector formulation of two-stage L2D, where the AI predictor model is fixed and separate from the trainable routing or rejector model, on Collab-CXR, a multilabel chest X-ray dataset with multiple human annotations per case. This is the first work to look at L2D in the context of real-world medical imaging data with human annotations. We further introduce a new setup, L2D with Guidance, where the decision space is extended to three choices: predict autonomously, defer to a human expert, or defer to a human expert and provide AI guidance. We compare multiple rejector architectures and loss functions, and different input feature availabilities. This is reproduced on two larger datasets, VinDr-CXR and CheXpert. Our results show that two-stage L2D with Guidance outperforms classic two-stage learning to defer, as well as human-alone, AI-alone and AI-guided human baselines. Notably, this performance is achieved with simpler loss functions compared to formally defined L2D surrogate loss functions in current literature.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-24

The authors' main claim, as the classifier picked it out 1.00

“Our results show that two-stage L2D with Guidance outperforms classic two-stage learning to defer, as well as human-alone, AI-alone and AI-guided human baselines.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.96 · states limitation 0.21 · human participants 0.67 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.05 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Medical image interpretation is high-volume and time-consuming, and while AI interpretation can reduce workload, fully autonomous deployment carries potential safety concerns and low specificity may in practice lead to increased clinician workload. background 1.00 0.05 0.02 hedged possibility 0.52
Learning to Defer (L2D) addresses this by selectively routing cases between autonomous prediction and human experts by learning from input features and AI model and human performance. background 0.90 0.06 0.04 asserted 1.00
While theoretical guarantees have been proven for L2D, its performance has not been validated on real-world medical datasets with human reader annotations. gap 1.00 0.05 0.02 asserted 1.00
We evaluate the predictor-rejector formulation of two-stage L2D, where the AI predictor model is fixed and separate from the trainable routing or rejector model, on Collab-CXR, a multilabel chest X-ray dataset with multiple human annotations per case. method 0.92 0.94 0.03 asserted 0.94
This is the first work to look at L2D in the context of real-world medical imaging data with human annotations. contribution 0.93 0.67 0.04 asserted 1.00
We further introduce a new setup, L2D with Guidance, where the decision space is extended to three choices: predict autonomously, defer to a human expert, or defer to a human expert and provide AI guidance. contribution 0.86 0.93 0.04 asserted 1.00
We compare multiple rejector architectures and loss functions, and different input feature availabilities. method 0.95 0.90 0.02 asserted 0.59
This is reproduced on two larger datasets, VinDr-CXR and CheXpert. method 0.81 0.85 0.04 asserted 0.96
Our results show that two-stage L2D with Guidance outperforms classic two-stage learning to defer, as well as human-alone, AI-alone and AI-guided human baselines. result 1.00 0.97 0.87 asserted 1.00
Notably, this performance is achieved with simpler loss functions compared to formally defined L2D surrogate loss functions in current literature. result 0.76 0.90 0.17 asserted 1.00

What kind of paper the classifier reports this is: method or tool 0.56 · evidence from observation or experiment 0.61

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-24: 1 request, 7,640 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (4)
Figure 1 of the paper
Figure 1. L2D with Guidance setup, where the rejector can choose to output autonomous AI prediction, defer to the AI-guided human expert, or defer to the plain, unguided human expert. Adapted from [12]. Human-AI Collaboration (HAIC), in particular Learning to Defer (L2D) [8], in healthcare offers a potential approach to support workload reduction. L2D is a supervised learning framework in which a rejector model learns, for each input, whether to output an autonomous decision from the AI predictor, or defer to an external expert, so as to minimise the overall expected system-level cost that…
Figure 2 of the paper
Figure 2. Accuracy of the AI predictor (yellow), non-guided human expert (blue) and AI-guided human expert (green) across pathologies. The AI predictor and AI-guided human outperform the non-guided human by the largest margin on pleural effusion, cardiomegaly and fracture. CheXpert and VinDr-CXR To validate the findings on Collab-CXR, the ex￾periments were repeated with two larger datasets, CheXpert [5] and VinDr-CXR [9]. While these datasets have individual real human annotations, removing the need for pooling, they do not contain human annotations with AI assistance. Fol￾lowing the standard…
Figure 3 of the paper
Figure 3. Performance of L2D with Guidance using (a) tabular features only, (b) image features, and (c) label-specific image embeddings, on Collab-CXR. Results are shown across different rejector models and input feature combinations, averaged over n = 20 random seeds, and compared to baselines. Note that the y-axis is truncated for legibility. Classic L2D Performance This new formulation significantly outperforms classic L2D on Collab-CXR (paired t-test, p < 0.0001) (
Figure 4 of the paper
Figure 4. Accuracy of the L2D system with MLP rejector, comparing L2D-G to classic L2D, with and without pathology-specific image embeddings, on Collab-CXR. L2D-G outperforms baselines of AI-guided human, AI alone and human alone. Classic L2D performs worse than baselines on this data. Note that the y-axis is truncated for legi￾bility. that with this moderately-sized dataset, without the option to defer to an AI￾guided human, there is model overfitting and the system cannot learn useful features to defer between the AI and a single choice of expert. Results on CheXpert and VinDr-CXR The same…

Figures are the authors', reproduced from the paper.

References (13)
  1. [1] Breiman, L.: Random forests. Machine Learning45(1), 5–32 (Oct 2001). https://doi.org/10.1023/a:1010933404324 on Pithdoi:10.1023/a:1010933404324
  2. [2] Chen,T.,Guestrin,C.:Xgboost:Ascalabletreeboostingsystem.arXiv:1603.02754 (2016) on PithPith's reviewarXiv
  3. [3] Commission, E., for Health, D.G., Safety, F., EEIG, Evidence, O., PwC: Study on the deployment of AI in healthcare – Final report. Publications Office of the European Union (2025). https://doi.org/doi/10.2875/2169577 on Pithdoi:10.2875/2169577
  4. [4] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. arXiv:1512.03385 (2015) on PithPith's reviewarXiv
  5. [5] Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al.: Chexpert: A large chest ra- diograph dataset with uncertainty labels and expert comparison. Proceedings of the AAAI Conference on Artificial Inte... on Pithdoi:10.1609/aaai.v33i01.3301590
  6. [6] Mao, A., Mohri, C., Mohri, M., Zhong, Y.: Two-stage learning to defer with mul- tiple experts. Advances in NeurIPS36, 3578–3606 (2023) on Pith
  7. [7] Moehring, A., Kutwal, M., Huang, R., Banerjee, O., Jacobi, A., Eber, C., Men- doza, D., Chung, M., Dayan, E., Gupta, Y., et al.: A dataset for understanding radiologist-artificial intelligence collaboration. Scientific Data12(1) (May 2025). https://doi.org/10.1038/s41597-025-05054-0 on Pithdoi:10.1038/s41597-025-05054-0
  8. [8] Mozannar, H., Sontag, D.: Consistent estimators for learning to defer to an expert. In: ICML. pp. 7076–7087. PMLR (2020) on Pith
  9. [9] Nguyen, H.Q., Pham, H.H., Tuan Linh, L., Dao, M., Khanh, L.: VinDr-CXR: An open dataset of chest X-rays with radiologist annotations. PhysioNet (Jun 2021). https://doi.org/10.13026/3akn-b287, https://doi.org/10.13026/3akn-b287, version 1.0.0 on Pithdoi:10.13026/3akn-b287
  10. [10] Rudolph, J., Huemmer, C., Preuhs, A., Buizza, G., Hoppe, B.F., Dinkel, J., Koliogiannis, V., Fink, N., Goller, S.S., Schwarze, V., et al.: Nonradiology health care professionals significantly benefit from AI assistance in emergency- related chest radiography interpretation. CH... on Pithdoi:10.1016/j.chest.2024.01.039
  11. [11] Smit, A., Jain, S., Rajpurkar, P., Pareek, A., Ng, A.Y., Lungren, M.P.: Chexbert: Combining automatic labelers and expert annotations for accurate radiology report labeling using bert. arXiv:2004.09167 (2020) on PithPith's reviewarXiv
  12. [12] Strong, J., Sun, E., Rogers, H., Higham, H., Noble, A.: Learning to defer: A survey. 10.5281/zenodo.17843044 (2025) on Pithdoi:10.5281/zenodo.17843044
  13. [13] Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classi- fication and localization of common thorax diseases. In: Proceedings of the IEEE Conference on Computer Vision and Patte... on Pith
Cited by 1 Pith paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment