PCXR-Copilot: Device-aware pediatric chest radiographic analysis with segmentation, suppression, and edema-related cues
Authors: Xueyang Li, Alec Girten, Peng Zeng, Mingze Jiang, Shaoyang Huang, Gelei Xu, Danny Z. Chen, Yiyu Shi, John N. Kheir
Organizations: Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN, USA. · Department of Cardiology, Boston Children’s Hospital, Boston, MA, USA. · Department of Pediatrics, Harvard Medical School, Boston, MA, USA. · School of Computing and Information, University of Pittsburgh, Pittsburgh, PA, USA.
Patients admitted to pediatric ICU and cardiac ICU often require intensive perioperative support and frequent chest radiography to monitor support devices, extravascular lung water, and body-wall edema. Radiographic interpretation and analysis is often difficult due to superimposed anatomic and iatrogenic structures that obscure clinically relevant signals. While deep learning can provide decision support, translation to pediatric radiographs is still limited by high device burden and scarcity of pediatric datasets with clinically grounded device annotations. To address this gap, we first introduce PediaDeviceSeg-1K, a curated dataset of 1,000 anterior-posterior radiographs of pediatric CICU patients with clinically grounded, fine-grained annotations of common ICU devices. To our best knowledge, this is the first publicly available pediatric CICU CXR resource with clinically grounded device annotations for device-focused benchmarking. Building on PediaDeviceSeg-1K, we further present PCXR-Copilot, a device-aware pipeline that outputs lung and device masks, optional device-suppressed views, and edema-related visualization maps and scalar indices. The full pipeline has been integrated and runs in real time within the TWIST platform at Boston Children's Hospital, demonstrating feasibility for real-world clinical deployment.
Figures & tables
Label
n
Label
n
Sternal Wire
732
End tidal CO2 monitor
72
Nasogastric Tube
578
Vent Tubing
61
EKG Lead
536
Tourniquet clip
56
EKG sticker
489
Left internal jugular vein catheter
48
Temporary pacing wires
427
Esophageal temperature probe
41
Endotracheal tube (ETT)
423
Cardiac tourniquet
35
Table 1: Medical device label prevalence in the annotated pediatric CXR dataset. The counts indicate the number of labeled images (out of 1,000) in which each label appears at least once (unique per image) after label mapping and casing unification.
Descriptor
Overall
Across folds (mean ± SD)
Patients, n
100
20.0 ± 0.0
Images, n
1000
200 ± 0.9
Images per patient
6.0 [3.0–14.0]
5.8 ± 0.3
Device categories per image
6.0 [4.0–9.0]
6.4 ± 0.5
Device instances per image
19.0 [11.0–29.0]
18.8 ± 1.3
Device pixel ratio (%)
2.70 [1.34–4.51]
2.67 ± 0.28
Table 2: Dataset complexity and device-burden descriptors. Summary statistics are computed from the labeled set (1,000 images). When applicable, values are reported as median [IQR] or mean ± SD. Device pixel ratio is the percentage of image pixels occupied by devices; device-in-lung ratio is the percentage of image pixels occupied by devices within the lung masks; device-lung overlap is the percentage of lung pixels occupied by devices.
Task
Model
Dice (%)
Binary (device vs. background)
nnU-Net
77.8±1.8
GLCP
77.5±1.8
U-Net++ (EfficientNet-b0 encoded)
76.2±1.6
U-Net++ (ResNet-18 encoded)
75.7±1.8
DeepLabv3+
74.8±1.4
Multi-label (prevalence ≥30% )
nnU-Net
68.2±2.7
Table 3: Medical device segmentation results. Binary segmentation is reported as Dice. Multi-label segmentation is evaluated on device categories with prevalence ≥30% ; Dice is macro-averaged over included labels. Values are reported as mean ± SD across five folds.
Figure 1: PCXR-Copilot end-to-end pipeline and representative outputs. (a) Overview of the PCXR-Copilot pipeline, showing the example input pediatric chest radiograph (PCXR) and the two analysis branches for pulmonary edema (PE) and body-wall edema (BWE). (b) Predicted lung masks overlaid on the input radiograph. (c) Expert device annotations (ground truth) for the input radiograph. (d) Device masks predicted by nnU-Net. (e) Device-suppressed image produced by mask-guided inpainting. (f) Pulmonary edema highlighting with the derived PE index. (g) Body-wall edema highlighting with the derived width-based index.
Figure 2: Lung water overlay in the TWIST radiology viewer. Screenshots of the TWIST radiology viewer show comparative chest radiographs of the same patient separated by 28 hours, in chronological order from left to right. (a) Two radiographs displayed without overlay. (b) The same two radiographs with an overlay depicting lung water, which facilitates recognition of the decrease in lung water between the two radiographs. Note the handling of the PICC line in the right upper extremity and the pilot balloon, both of which are subtracted from the lung water overlay computation.
Figure 3: Representative medical device suppression and associated image changes. For each case, the original radiograph (left), device-suppressed radiograph (middle), and pixel-wise difference map (right) are shown. Difference maps highlight regions altered by suppression, with changes concentrated predominantly along device-containing regions and limited changes in surrounding non-device anatomy.
Figure 4: Within-patient evolution of the pulmonary edema index. Index values for 15 frontal radiographs acquired during a single admission; numbered markers denote the four timestamps shown below. Lung-water overlays are restricted to the predicted lung fields and share a common color scale across the panels.
Figure 5: Device-related dependence and external evaluation of the PE index. ( a,b ) Rank–rank plots show the association between device–lung overlap and the PE index in the BCH cohort with devices retained (a) and suppressed (b). Solid lines show least-squares fits to the ranked values; Spearman ρ decreased from 0.126 to 0.049 after suppression. ( c ) Distribution of the lung-field intensity index across edema grades in MIMIC-CXR, using the BCH-trained lung segmentation model: grade 0, no edema ( n=279 ); grade 1, vascular congestion ( n=153 ); grade 2, interstitial edema ( n=134 ); and grade 3, alveolar edema ( n=82 ). Boxes show medians and interquartile ranges; whiskers extend to the most extreme observations within 1.5 interquartile ranges, and points represent individual radiographs. ( d ) ROC curve for grade 3 versus grade 0, based on 361 radiographs from 170 patients. The 95% confidence intervals for Spearman ρ and AUC were obtained from 2,000 patient-level bootstrap resamples.
Figure 6: Qualitative visualization of body-wall edema region segmentation and post-processing. (a) Representative chest radiographs showing the original image, manual ground-truth annotation (yellow overlay), and model prediction (magenta overlay). The predicted panel reports the Dice score against ground truth and the left, right, and average body-wall width estimates (in mm). (b) A post-processing comparison example, showing ground truth (yellow), the initial prediction before post-processing, and the post-processed prediction after retaining the two longest connected components and removing small spurious regions. The post-processing reduces isolated false positives and improves overlap with ground truth, as reflected by the Dice score.
Figure 7: Cohort-level distribution of predicted body-wall edema thickness. Violin and box plots summarize the radiograph-level distribution of predicted body-wall edema thickness (in mm) for the body-left and body-right regions and their average. Points are horizontally jittered within each category for visualization.
Background and Objective: External evaluation of medical-imaging AI is often collapsed into discrimination. We evaluated a computational protocol that separately tests discrimination, probability calibration, fixed operatingpoint transport, shortcut-associated signal, and limited-label recoverability for pediatric pneumonia classification across datasets from three countries. Methods: After exact-duplicate removal, 5,824 Guangzhou radiographs supported leakage-controlled source development and internal testing. A frozen three-seed DenseNet121 dual-view ensemble was evaluated zero-shot on BDCXR-3257 from Bangladesh (n = 3, 257) and an untouched harmonized VinDr-PCXR/PediCXR test cohort from Vietnam (n = 1, 077). Matched seed-42 variants tested architectural robustness. Secondary BDCXR analyses used a fixed 651-image adaptation pool and 2,606-image hold-out; 163, 326, and 651 labels represented 5%, 10%, and 20% of complete BDCXR. Results: Internal AUROC was 0.976 with 95.1% sensitivity. BDCXR and VinDr-PCXR AUROC were 0.798 and 0.742, while frozen-threshold sensitivity fell to 6.2% and 0%. Source-to-BDCXR AUROC degradation occurred for a full-image baseline (0.961 to 0.749), ungated dual-view model (0.977 to 0.766), and gated MixStyle model (0.966 to 0.789). With 163 BDCXR labels, Platt recalibration preserved AUROC while increasing held-out sensitivity to 88.3%, but specificity was 47.9% and the alert rate was 78.5%. Two hundred repeated 163-label fits confirmed sensitivity recovery but substantial specificity variability. Conclusions: Cross-dataset shifts across countries affected ranking, probability alignment, and source-defined decision behavior differently. Transport studies should evaluate these components separately and quantify the operational burden of apparent recovery.
Nazim-E-Alam
Department of Computer Science, American International University-Bangladesh (AIUB), 408/1 (Old KA 66/1), Kuratoli, Khilkhet, Dhaka 1229, Bangladesh
Clinical diagnosis often requires combining imaging with physiological measurements, yet deployed models typically operate on unimodal data. We present PaCX-MAE, a cross-modal distillation framework that injects physiological priors into chest X-ray (CXR) encoders while remaining strictly unimodal at inference. PaCX-MAE augments in-domain masked autoencoding with a dual contrastive-predictive objective, aligning CXR representations with paired ECG and laboratory embeddings. Extensive evaluation across nine benchmarks demonstrates consistent improvements over domain-specific MAE, particularly on physiology-dependent tasks (e.g., +2.7 AUROC on MedMod; +6.5 F1 on VinDr). The method proves highly label-efficient in the 1% regime and preserves anatomical fidelity, achieving parity with MAE on segmentation tasks. Zero-shot and attention analyses confirm that PaCX-MAE successfully learns to attend to physiological indicators, such as the cardiac silhouette, absent in standard visual pretraining.
Yancheng Liu, Kenichi Maeda, Manan Pancholy
Department of Computer Science, Brown University, Providence, RI, USA.
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.
Md Shihabul Islam Shovo, Nishi Kanta Paul, Jamil Fayyad
NIMISHES Lab, Dhaka, Bangladesh · American University of Sharjah, Sharjah, UAE