AT-CXR: Uncertainty-Aware CXR Triage System
- The paper introduces an agentic triage framework for pulmonary-edema that integrates uncertainty checks and a three-outcome decision process (accept, escalate, or defer) to enhance diagnostic safety.
- It employs a multi-stage system combining deterministic rule-based and LLM routers to balance throughput with improved accuracy and low-risk selective prediction.
- The framework demonstrates superior risk-coverage metrics and latency benefits compared to conventional CXR classifiers, advancing automated chest radiograph interpretation.
AT-CXR most explicitly denotes the uncertainty-aware agentic triage framework introduced for chest X-rays in 2025, where automated interpretation is framed as a sequential decision problem that must decide whether to accept a case, escalate analysis through additional tools, or abstain and defer to a human reader under explicit uncertainty and latency constraints (Li et al., 26 Aug 2025). In adjacent chest-radiography literature, the same label or closely related “AT-CXR-style” usage also appears in a broader sense for automated CXR systems that quantify disease burden, segment anatomy, classify thoracic pathology, generate reports, or support anatomy-centric reasoning (Jr. et al., 2020, Kamal et al., 2021, Li et al., 5 May 2025). This dual usage matters because it distinguishes a specific agentic triage architecture from the wider research program of automated chest X-ray interpretation.
1. Terminology and problem formulation
In its most specific usage, AT-CXR is an end-to-end framework for pulmonary-edema triage on chest radiographs. Its defining claim is not merely that a model predicts edema, but that triage should be treated as a policy over actions with explicit guardrails for uncertainty, out-of-distribution detection, latency, throughput, and auditable human fallback (Li et al., 26 Aug 2025).
The core action space has three operational outcomes. A case may be accepted automatically, escalated through additional verification tools, or abstained with handoff to a human together with a suggested label and rationale. This differs from conventional one-shot classification, where a system typically emits a probability or binary label without governing analysis depth or deferral. A common misconception is therefore to view AT-CXR as just another edema classifier; the 2025 formulation is instead a constrained decision process over diagnostic actions (Li et al., 26 Aug 2025).
The broader literature shows why this framing emerged. Earlier CXR automation emphasized task-specific outputs such as disease scoring, segmentation, or report generation, often with strong performance on narrowly defined tasks but without an explicit stop/escalate/defer mechanism. This suggests that AT-CXR, in its agentic sense, is best understood as a later systems-level layer built on top of earlier perception and reasoning capabilities rather than a replacement for them (Jr. et al., 2020, Zhang et al., 2020).
2. Architecture of the uncertainty-aware triage system
The 2025 AT-CXR system is organized into four stages: data ingestion, uncertainty checking, agentic routing, and generation of triage and explainability artifacts (Li et al., 26 Aug 2025).
Stage I: Data ingestion is automated through a folder watcher. Supported inputs are DICOM, JPEG, and PNG. Preprocessing includes DICOM-to-PNG conversion, PHI removal from headers when present, grayscale conversion, resizing to , and normalization to . This design is explicitly aimed at workstation-compatible deployment rather than batch-only offline evaluation (Li et al., 26 Aug 2025).
Stage II: Uncertainty check computes two per-case signals. The first is a baseline edema posterior from a supervised classifier, RexNet-150, represented as . The second is a Mahalanobis OOD score in standardized radiomics feature space derived from PyRadiomics features extracted on a whole-image mask with a one-pixel zero border. The score for a new case is defined as
This paired uncertainty representation is central: confidence alone is not treated as sufficient because a high posterior on an OOD case is still considered unsafe (Li et al., 26 Aug 2025).
Stage III: agentic decision routing is governed by an uncertainty guardrail. If a case is OOD or low-confidence, direct acceptance is forbidden; the router must invoke verification tools, escalate, or abstain. Two router designs are evaluated: a deterministic rule-based router and an LLM-decided router. They share the same inputs, thresholds, actions, and tools, differing only in the policy used to choose the next action (Li et al., 26 Aug 2025).
The tool library is ordered by cost as
TTA applies benign augmentations and measures posterior mean and variance. MoE uses a four-model committee comprising RexNet-150, RegNetY-1.6, ResNet-50, and EfficientNet-V2, with majority-vote agreement rate . VLM escalation uses GPT-4.1-mini with vision and is terminal: it produces abstention with suggested label and rationale rather than automated acceptance. The thresholds used in experiments are , as the 95th percentile of reference Mahalanobis scores, , and 0 (Li et al., 26 Aug 2025).
Stage IV: triage and explainability artifacts writes accepted cases into positive or negative folders and sends abstained cases to a Human_Intervention_Needed queue. Accepted positive cases receive a class activation map and a lung whiteness index. After rib suppression and device suppression, the lung whiteness index is
1
with higher values interpreted as greater parenchymal opacity. Every case also receives a machine-readable JSON trace containing case ID, timestamps, versions, thresholds, uncertainty scores, tool sequence, final decision, and deciding agent (Li et al., 26 Aug 2025).
3. Performance, selective prediction, and deployment trade-offs
AT-CXR was evaluated on a balanced subset of NIH ChestX-ray14 containing 1,000 frontal radiographs, with 500 pulmonary-edema positive and 500 pulmonary-edema negative cases, using five-fold stratified cross-validation and a single NVIDIA A10 GPU (Li et al., 26 Aug 2025).
At full coverage, the LLM router achieved 95.3 ± 3.3 accuracy, 96.7 ± 4.1 precision, and 93.8 ± 2.8 recall. The rule-based router achieved 93.8 ± 3.9 accuracy, 96.0 ± 3.6 precision, and 92.9 ± 3.6 recall. The best non-AT-CXR supervised baseline listed was EfficientNet-V2 at 87.8 ± 3.0 accuracy, while the best zero-shot VLM listed was MedGemma-27B at 81.2 ± 3.9 accuracy (Li et al., 26 Aug 2025).
The more distinctive result is on selective prediction, where AT-CXR is evaluated by area under the risk-coverage curve, risk at 80% coverage, and coverage at 5% risk. The LLM router achieved AURC 0.9 ± 1.1, risk@80% coverage 1.4 ± 1.7, and coverage@5% risk 98.9 ± 2.5. The rule-based router achieved AURC 1.0 ± 1.1, risk@80% coverage 1.5 ± 1.7, and coverage@5% risk 98.5 ± 2.2. By comparison, the supervised baselines had AURC values around 4.0–5.1, risk@80% coverage around 6.5–9.1, and coverage@5% risk around 59–69%. These numbers show that the framework is optimized not only for accuracy but for low-risk automation at high coverage (Li et al., 26 Aug 2025).
Latency is similarly part of the design target. Mean per-image latency was 1.5 s for the rule-based router and 3.3 s for the LLM router, compared with 6.6 s for GPT-4o, 5.7 s for GPT-4.1, 4.6 s for GPT-4.1-mini, 20.5 s for GPT-5, 15.2 s for GPT-5-mini, 7.9 s for MedGemma-4B, 885.1 s for MedGemma-27B, and 3.8 s for CheXagent. The paper therefore presents the two routers as complementary operating points: the rule-based version prioritizes deterministic behavior and throughput, while the LLM router prioritizes maximal accuracy and slightly better selective-prediction behavior (Li et al., 26 Aug 2025).
4. Pre-agentic foundations: CAD, CT-grounded supervision, and anatomy priors
Long before AT-CXR was framed as an agentic policy, chest-radiograph automation had already developed strong task-specific foundations. A major line was computer-aided triage. CAD4TB v6, a deep-learning commercial system for pulmonary tuberculosis screening, processed a chest radiograph in less than 15 seconds, output a 0–100 score, and achieved 76% specificity at 90% sensitivity against Xpert MTB/RIF on an independent set of 5565 images, while reaching 98% specificity at 90% sensitivity against a five-expert radiological reference standard (Murphy et al., 2019). It also showed concrete programmatic effects: at 90% sensitivity, users could process 132 subjects/day at \$[0,1]$28.38 for version 3. This established that automated CXR systems could already function as operational triage tools, although without the explicit uncertainty-routing machinery later emphasized in AT-CXR (Murphy et al., 2019).
A second foundation was CT-grounded severity quantification. For COVID-19 airspace disease, one system trained a CNN on digitally reconstructed radiographs derived from CT-based ground truth, using volumetric percentage of opacity on CT ($[0,1]$3) and area-based percentage of opacity on radiographs ($[0,1]$4) as linked severity measures. In a cohort of 86 COVID-19 patients with positive RT-PCR and paired CT and CXR within 48 hrs, the ensemble CNN on CXR achieved MAE 9.56% and Pearson $[0,1]$5 against CT ground truth, compared with expert-reader average MAE 11.98% and $[0,1]$6, with runtime 52 ms per CXR on an RTX 2080 Ti GPU (Jr. et al., 2020). This showed that CT-derived supervision could convert chest radiography from a subjective scoring problem into a quantitative biomarker estimation problem.
A third foundation was anatomy-aware perception. Anatomy-XNet introduced a DenseNet-121-based classifier with Anatomy Aware Attention and Probabilistic Weighted Average Pooling, using semi-supervised segmentation to transfer lung and heart masks from JSRT to large unannotated datasets. It reported AUC scores of 85.78% on NIH, 92.07% on CheXpert, and 84.04% on MIMIC-CXR, while degrading by only about 0.2% under artificially cut masks, unlike competing mask-based methods that degraded by around 2.41–3.13% (Kamal et al., 2021). The key idea was that radiographic classification should first localize anatomy, then assess abnormality.
That anatomy-centric line was later scaled in synthetic-supervision form by AnyCXR, which trained entirely on CT-derived DRRs using Multi-stage Domain Randomization and Conditional Joint Annotation Regularization. The framework produced over 100,000 synthetic CXRs from 2,982 curated CT scans, segmented 54 anatomical structures, and reported mean DSC 92.5% and HD 10.71 px in PA view and mean DSC 88.8% and HD 20.41 px in LA view on zero-shot TotalSegmentator testing (Zifei et al., 19 Dec 2025). A plausible implication is that the perceptual substrate needed by agentic systems such as AT-CXR increasingly relies on synthetic supervision, anatomy priors, and cross-modal supervision rather than image-level labels alone.
5. Extension from perception to reasoning and report generation
AT-CXR-style automation also encompasses systems that move beyond detection into language generation, structured reasoning, and multi-agent orchestration. An early weakly supervised report-generation system trained on 12,219 chest X-ray images and 12,219 paired radiological reports from Nanjing Drum Tower Hospital used DenseNet-121, an LSTM, and soft attention to learn directly from image-report pairs without extra disease labels. It achieved an average CIDEr of about 5.8, and in blind review both human and model reports were rated acceptable at 86% when scores of 4 or above were counted as acceptable (Zhang et al., 2020).
Later work shifted the evaluation target from lexical overlap to clinical acceptability. CXRMate-2, trained with structured multimodal conditioning and reinforcement learning, improved GREEN by 11.2% and RadGraph-XL by 24.4% on MIMIC-CXR relative to MedGemma 1.5 (4B), and in a blinded retrospective study with three consultant radiologists, generated reports were deemed acceptable in 45% of ratings. Preference for radiologist reports was driven primarily by higher recall, whereas generated reports were often preferred for readability (Nicolson et al., 21 Apr 2026). This clarifies that automated CXR interpretation is increasingly evaluated as a clinically constrained communication task rather than a captioning task.
Reasoning-centered systems push further. AOR, or Anatomical Ontology-Guided Reasoning, centers multimodal reasoning on cross-modal region-level information comprising region name, coordinates, and ROI visual representation. It organizes 38 anatomical regions, 68 attributes, and object-attribute restriction relations into physician-guided ontology-based chain-of-thought supervision, with 290k VQA samples and 532k report-generation pairs in AOR-Instruction. On MIMIC-CXR-VQA, AOR-t achieved 80.48 on verify, 71.96 on choose, and 65.05 on query, while report generation on MIMIC-CXR reached ROUGE-L 25.37, BERTScore 83.92, and F1CheXbert 46.53 (Li et al., 5 May 2025).
Agentic orchestration generalizes this reasoning structure to tool use. CXRAgent uses a director to coordinate tool invocation, an Evidence-driven Validator to ground tool outputs in visual evidence, diagnostic planning through collaboration modes such as Skip, Relay, Dispatch, and Probe, and collaborative decision-making with recruited specialists. It reached 67.0% overall accuracy on CheXbench, 75.6% overall accuracy on Medical-CXR-VQA, and RaTEScore 0.513 on MIMIC-CXR report generation (Lou et al., 24 Oct 2025). Relative to such systems, AT-CXR’s distinctive contribution is narrower but operationally sharp: it focuses on the triage policy itself, with uncertainty guardrails and selective automation as first-class objects.
6. Limitations, misconceptions, and likely trajectory
The 2025 AT-CXR formulation is limited to pulmonary-edema triage on a balanced subset of a single dataset, and the authors explicitly identify future extensions to multi-label CXR triage and more rigorous evaluation of medical-device suppression by feeding suppressed images to the router directly (Li et al., 26 Aug 2025). This means that current claims should not be read as establishing general-purpose autonomous chest-radiograph triage across the full clinical label space.
A second limitation concerns synthetic-to-real transfer. In a study of coronary artery calcification detection using synthetic chest X-rays generated from ECG-gated cardiac CT, the best DRR-based configuration achieved mean AUC 0.754 ± 0.055, competitive with prior CXR-based studies, but transfer from DRRs to actual CXRs was not evaluated. The paper therefore described its result as feasibility rather than clinical deployment and identified domain adaptation to real CXRs as an important next step (Saeed et al., 14 Nov 2025). This caution applies broadly to AT-CXR-style systems that rely on CT-derived supervision.
A third limitation is modality mismatch in severity estimation. In COVID-19 monitoring, a CXR-based “Pneumonia Ratio” pipeline validated against DRRs synthesized from serial CT scans showed correlation 0.74 with CT ratios and trend agreement accuracy 0.87 in 9 severe COVID-19 patients, while explicitly noting that CXR is less sensitive than CT for mild disease and is more reliable for moderate-to-severe pneumonia monitoring (Amer et al., 2020). This reinforces a recurring point: radiographic automation can be quantitatively strong without being equivalent to CT across all disease regimes.
The larger trajectory is nevertheless clear. Automated CXR systems have progressed from scoring engines and single-task CAD to CT-grounded quantification, anatomy-aware segmentation and classification, clinically conditioned multimodal models, structured reasoning systems, report generators evaluated by radiologists, and finally uncertainty-aware agents that treat triage as a decision policy rather than a raw prediction problem (Murphy et al., 2019, Jr. et al., 2020, Zifei et al., 19 Dec 2025, Nicolson et al., 21 Apr 2026, Li et al., 26 Aug 2025). This suggests that the long-term meaning of AT-CXR is not merely automation of chest X-ray interpretation, but automation under explicit clinical constraints: uncertainty, abstention, evidence, workflow compatibility, and auditable human fallback.