PaperlayerView exact revision on arXiv

Source-linked AI summary

ZAGNet: Zone-Aware Graph Aggregation Network for Patient-Level Lung Ultrasound Diagnosis

Li Chen, Shubham Patil, Rashid Al Mukaddim, Jochen Kruecker, Balasundar Raju, Alvin Chen

arXiv:2610.02263v1eess.IVcs.CV

TL;DR

Patient-level lung ultrasound diagnosis must combine findings across zones, but common frame- or loop-based methods use pooling that does not model inter-zone relationships. ZAGNet graphs temporally tracked findings by anatomical adjacency and uses attention-based reasoning for patient-level prediction. It outperforms pooling-based aggregation for consolidation and pleural effusion in the reported evaluation.

  • Problem

    Existing lung ultrasound methods often analyze individual frames or loops and use pooling that ignores anatomical relationships, while patient-level diagnosis depends on findings across zones.

    Source: §I · p. 1

  • Method

    ZAGNet represents temporally tracked findings as graph nodes connected by anatomical zone relationships, applies graph message passing, and predicts from the graph using patient-level labels.

    Source: §I · p. 1, D. Graph Reasoning · p. 3

  • Results

    ZAGNet achieved consolidation AUC 0.803 versus 0.677 (max pooling) and 0.674 (mean pooling), and pleural effusion AUC 0.893 versus 0.804 (max pooling) and 0.815 (mean pooling).

    Source: D. Results · p. 3

  • Takeaways & Limitations

    The results support anatomical inter-zone reasoning as a more informative approach than independent regional aggregation for patient-level diagnosis, including exams with variable or incomplete scanning protocols.

    Source: D. Results · p. 3, D. Graph Reasoning · p. 3

Abstract

from arXiv · show

Patient-level lung ultrasound (LUS) diagnosis requires integrating findings acquired across multiple anatomical zones, yet clinical examinations frequently involve variable and incomplete scanning protocols with missing zones. Existing diagnostic AI methods primarily analyze individual frames or video loops, relying on heuristic aggregation strategies such as max or mean pooling that ignore inter-zone relationships for patient-level inference. This paper presents ZAGNet, a Zone- Aware Graph Neural Network that represents temporally tracked pathology findings as graph nodes connected by anatomical zone adjacency. A graph transformer network propagates contextual information across neighboring lung regions, while a virtual global node aggregates graph-level features to predict patientlevel consolidation and pleural effusion using only patient-level supervision. ZAGNet accommodates missing zones by computing on a graph structure without fixed input format or size. We evaluate ZAGNet on a multicenter dataset of 714 subjects (20,256 LUS video loops) with exams varying from 4 to 16 zones across anterior, posterior, and lateral thoracic regions. For consolidation diagnosis, ZAGNet achieved an AUC of 0.803 compared to 0.677 (max pooling) and 0.674 (mean pooling). For pleural effusion, AUC increased to 0.893 from 0.804 (max pooling) and 0.815 (mean pooling). These represent improvements of up to 19% and 11% for consolidation and pleural effusion respectively. The results demonstrate that graph-based inter-zone reasoning provides an effective and clinically consistent framework for automated patient-level LUS assessment.

I. INTRODUCTION

Patient-level lung ultrasound diagnosis must integrate findings across anatomically related zones, but existing approaches often aggregate regional predictions without modeling those relationships. ZAGNet represents tracked findings in an anatomical graph and outperforms pooling-based aggregation for both target diagnoses.

  • Clinical patient-level assessment depends on both pathology within each zone and its spatial distribution across the lung.Source: §I · p. 1
  • Existing methods often use max or mean pooling, which ignores anatomical relationships and is not designed for exams with missing or variable numbers of zones.Source: §I · p. 1
  • ZAGNet builds nodes from temporally tracked pathology findings, connects them by anatomical adjacency, and uses a virtual global node for patient-level prediction under weak supervision.Source: §I · p. 1
  • ZAGNet outperforms conventional pooling-based aggregation for both consolidation and pleural effusion classification.Source: §I · p. 2

A. Problem Formulation

ZAGNet formulates patient-level diagnosis as graph-level classification from detected findings across a patient's lung ultrasound loops. It represents each temporally consistent pathology tracklet as a node with features describing its location, extent, persistence, and class.

  • Each patient-specific graph is built from detected pathological findings to predict a binary patient-level label for consolidation or pleural effusion.Source: A. Problem Formulation · p. 2
  • ZAGNet associates consecutive detections into tracklets and discards tracklets shorter than five frames to suppress unstable detections and reduce false positives.Source: A. Problem Formulation · p. 2
  • Tracklets, rather than scan zones, become separate nodes, allowing multiple independent findings in one zone to be modeled separately.Source: A. Problem Formulation · p. 2
  • Each node's 8-dimensional feature vector encodes its zone, normalized bounding box, mean detection confidence, normalized duration, and pathology class.Source: A. Problem Formulation · p. 2

C. Graph Construction

ZAGNet connects finding nodes according to within-zone co-occurrence and predefined anatomical adjacency between zones. The resulting graph reflects observed findings, supports missing zones, and uses a virtual global node to summarize the examination.

  • Within-zone edges connect all tracklets in a zone, while inter-zone edges connect nodes from anatomically adjacent zones in the predefined map.Source: C. Graph Construction · p. 2
  • Observed tracklets determine which nodes exist, so missing scan zones require neither padding nor imputation.Source: C. Graph Construction · p. 2
  • A virtual global node connects to every tracklet node and aggregates information from the whole examination into a context-aware representation.Source: C. Graph Construction · pp. 2–3

D. Graph Reasoning

Graph transformer layers use attention-based message passing to incorporate information from neighboring findings. The final virtual global-node embedding feeds a patient-level classifier trained using patient labels alone, while graph inputs accommodate variable and incomplete exams.

D. Graph Reasoning

  • Graph transformer convolution updates each node by aggregating neighbor features through attention-based message passing.Source: D. Graph Reasoning · p. 3
  • The final virtual global-node embedding summarizes pathological information across the examination and is used for patient-level prediction.Source: D. Graph Reasoning · p. 3
  • Training uses patient-level labels without zone-level supervision, and graph representations accommodate variable numbers of findings and incomplete scanning protocols.Source: D. Graph Reasoning · p. 3

III. EXPERIMENTS AND RESULTS

The multicenter evaluation included 714 subjects and 20,256 video loops, with examinations spanning 4–16 anatomical zones. Clinicians established patient-level ground truth, and data were split by subject.

III. EXPERIMENTS AND RESULTS

  • 714 subjects and 20,256 video loops from 11 clinical sites formed the multicenter emergency- and ICU-based dataset.Examinations covered 4–16 anterior, lateral, and posterior scan zones.Source: §III · p. 3
  • Clinicians established consolidation and pleural effusion ground truth using all available clinical information, including LUS, CT, laboratory tests, and medical history.Source: §III · p. 3
  • The dataset was divided into training, validation, and testing sets using subject-level partitioning.Source: §III · p. 3

B. Implementation Details

ZAGNet was implemented as a compact PyTorch model and trained with patient-level labels only.

B. Implementation Details

C. Experiments

The experiments compared graph reasoning with max- and mean-pooling aggregation while holding the underlying detection model constant, using AUC for evaluation.

C. Experiments

  • Max and mean pooling replaced the graph reasoning module to provide aggregation baselines with the same underlying detection model.Source: C. Experiments · p. 3
  • Performance was evaluated using the area under the receiver operating characteristic curve (AUC).Source: C. Experiments · p. 3
  • The baseline design was intended to isolate performance differences attributable to the aggregation strategy.Source: C. Experiments · p. 3

D. Results

ZAGNet outperformed both pooling baselines for consolidation and pleural effusion classification. Its graph message passing integrates neighboring-zone context and accommodates variable, incomplete scans.

D. Results

  • Graph message passing propagates context between neighboring regions, enabling more clinically consistent integration while accommodating variable and incomplete scanning protocols.Source: D. Results · p. 3
  • 0.803 AUC for consolidation exceeded 0.677 with max pooling and 0.674 with mean pooling, an improvement of up to 19%.Source: Table I · p. 3, D. Results · p. 3
  • 0.893 AUC for pleural effusion exceeded 0.804 with max pooling and 0.815 with mean pooling, an improvement of up to 11%.Source: Table I · p. 3, D. Results · p. 3

IV. CONCLUSION

ZAGNet models tracked pathological findings as graph nodes linked by anatomical zone relationships, enabling structured reasoning across scan regions and handling variable, incomplete exams with patient-level information. Multicenter experiments showed substantial improvements over pooling-based aggregation for consolidation and pleural effusion classification.

IV. CONCLUSION

  • Experiments on a multicenter dataset showed substantial improvements over conventional pooling-based aggregation for consolidation and pleural effusion classification.The conclusion does not report quantitative values.Source: §IV · p. 3
  • ZAGNet represents tracked pathological findings as graph nodes connected through anatomical zone relationships for structured reasoning across scan regions.Source: §IV · p. 3
  • The framework handles variable and incomplete exams using only patient-level information.Source: §IV · p. 3
  • Future work will investigate learned graph topologies, joint detection and aggregation training, additional pulmonary pathologies, and validation on more diverse clinical datasets.Source: §IV · p. 3
Loading 2610.02263v1…