OpenOrchestratorCT-Pilot: Coordinating Heterogeneous AI Modules for Automated Thoracic CT Analysis

Architecture overview

OpenOrchestratorCT-Pilot was designed following a runner–orchestrator architecture, in which a central pipeline coordinates multiple independent processing modules (“runners”), each encapsulating a distinct external analysis tool ( Figure 1). This design enables modular integration of heterogeneous algorithms within a unified workflow.

Figure 1. Architecture diagram.

Schematic of the OpenOrchestrator-CT pipeline. A NIfTI CT file serves as input to the orchestrator, which dispatches it to three analysis modules, each running in its own isolated conda environment via dedicated runner scripts that handle format conversion and execution. Each module produces a structured result, and the orchestrator aggregates these into a unified summary report.

Each module implements a standardized BaseRunner interface, consisting of two core methods: check_installation(), which verifies environment and dependency availability, and run(), which executes the analysis task. This abstraction ensures interoperability and facilitates the addition of new modules without modifying the core pipeline.

To prevent dependency conflicts between modules, each runner is executed within its own isolated Conda environment using conda run. A dedicated temporary working directory is created for each pipeline execution; within this directory, each module operates in a separate subfolder, ensuring clean input/output separation and enabling systematic cleanup after execution.

In this framework, OpenOrchestratorCT-Pilot does not embed or redistribute the external analysis modules it coordinates. Instead, users are responsible for independently installing and maintaining each module within its dedicated Conda environment. This design choice preserves modularity, respects upstream licensing constraints, and allows each component to evolve independently.

For this pilot implementation, three representative tasks were selected to demonstrate the technical feasibility of the framework: organ segmentation with volumetric abnormality detection, emphysema quantification, and pulmonary nodule detection. Importantly, this work is intended to illustrate the orchestration of heterogeneous tools rather than to assess the clinical performance or relevance of the individual algorithms.

Module 1: TotalSegmentator

The first module integrates TotalSegmentator for automated multi-organ segmentation from thoracic CT images in NIfTI format. TotalSegmentator is a deep learning–based framework that provides comprehensive anatomical segmentation and quantitative measurements.¹⁹

In this study, four structures were specifically analyzed: the heart, thyroid gland, and bilateral adrenal glands. These organs were selected based on their relevance for incidental findings in thoracic imaging. Segmentation was performed using the --fast configuration (3 mm model) along with the --statistics option, generating a structured statistics.json output. Extracted organ volumes (converted from mm³ to mL) were compared against normative reference values derived from recent literature.²⁰ Volumes exceeding predefined thresholds were flagged as potentially abnormal.

Module 2: Emphysema quantification

The second module performs emphysema quantification using an open-source tool based on density thresholding. Specifically, it computes the percentage of low-attenuation areas below −950 Hounsfield Units (LAA-950), a widely used quantitative biomarker of emphysema severity.²¹

Because the underlying implementation requires DICOM input, an intermediate conversion from NIfTI to DICOM is performed within the pipeline using the nibabel and pydicom libraries. Special care is taken to preserve spatial orientation and voxel geometry, including axis transpositions and flips to match clinical imaging conventions. To maintain strict module isolation, the pipeline dynamically generates and executes a temporary Python script within the module’s environment via conda run, avoiding any modification of the original repository.

A threshold of LAA-950 ≥ 10% was used to define significant emphysema. The module outputs the LAA-950 percentage (low-attenuation areas below −950 HU). A finding is flagged as significant emphysema when this value reaches or exceeds the 10% threshold.

Module 3: Lung nodule detection

The third module integrates a deep learning–based pulmonary nodule detection framework relying on a 3D Vision Transformer architecture (VitDet3D) trained on the LUNA16 dataset²² (available at: https://github.com/rlsn/LungNoduleDetection). The model processes volumetric CT data to identify candidate nodules using a sliding-window strategy (crop size 40 × 128 × 128, stride ratio 0.75) and outputs bounding box predictions with associated logit scores.

For each detected candidate, the module outputs world coordinates (in millimeters), estimated diameter, and confidence score derived from the logit threshold (default 0.0, corresponding to a predicted probability ≥50%). Nodules with a diameter ≥ 6 mm are flagged as clinically significant according to Fleischner Society guidelines (2017). All detections are compiled into a structured JSON output, which the orchestrator integrates into the final summary report.

Report generation

Following execution of all enabled modules, results are aggregated into a structured plain-text report generated by a dedicated reporting script ( Figure 2). The report summarizes module execution status, quantitative outputs (e.g., organ volumes with reference ranges), and all flagged findings based on predefined thresholds.

Figure 2. Example of a generated thoracic CT analysis report from a scan with pulmonary embolism from the ChestPathCT5-S100 database.

The report includes: (1) organ segmentation with measured volumes for heart, thyroid, and adrenal glands, (2) emphysema quantification expressed as the percentage of low-attenuation areas below −950 Hounsfield Units (LAA-950), and (3) lung nodule detection with coordinates and size of detected nodules. The structured output illustrates the type of summary provided by the automated pipeline.

Each report is automatically named using the input scan identifier and a timestamp to ensure traceability and reproducibility. A standardized disclaimer is included to emphasize the research-only nature of the tool.

Operation and system requirements

The pipeline was developed for Python 3.11 and requires Conda for environment management. It supports execution on systems equipped with NVIDIA GPUs (CUDA-enabled), and was tested on a Linux HPC environment.

Installation involves cloning the main repository, creating a dedicated orchestrator environment, and setting up individual Conda environments for each module. External module repositories are cloned into a designated modules/directory.

Pipeline configuration is centralized in a single config.yaml file, allowing users to enable or disable modules and define device allocation and file paths.

Dataset 1: ChestPathCT5-S100 (N = 87)

ChestPathCT5-S100²³ is an open, curated, anonymized institutional dataset comprising 87 thoracic CT scans representative of five common pathological categories: rib fracture (RF, N = 14), pleural fluid (N = 16), pulmonary mass (N = 17), pulmonary embolism (PE, N = 20), and pneumothorax (N = 20). The dataset includes both contrast-enhanced and non-contrast acquisitions, reflecting heterogeneous clinical acquisition conditions. All scans were processed in batch mode using a HPC batch script on the institutional HPC cluster (SLURM, 1 GPU per job, 16 cores, 128 GB RAM, maximum wall-clock time of 1 hour per job).

All 87 cases (100%) completed with full success across all three modules, with no partial or total failures observed in any pathology group ( Table 1). Total batch processing time was 268.0 minutes, corresponding to a mean of approximately 3.1 minutes per case.

Across the 87 processed cases, the three modules collectively produced the following outputs ( Table 2). TotalSegmentator flagged at least one pathological organ volume in 12.6% of cases overall, with the highest rate in the RF group (21.4%); mean heart volume was 519.33 ± 123.72 mL across the cohort. A high rate of undetected peripheral organs was observed (87.4% of cases with ≥1 undetected organ, mean 1.61 ± 0.94 per case), predominantly affecting the thyroid gland (63.2%) and adrenal glands, which is expected given the limited field of view of standard thoracic protocols. Emphysema quantification yielded a mean LAA-950 of 2.57 ± 6.08%, with 9.2% of cases exceeding the ≥10% significance threshold; the pneumothorax group showed the highest values (mean 7.94 ± 10.24%, 30.0% pathological rate), likely reflecting artifactual inclusion of intrapleural air by the density-thresholding algorithm, while the pleural fluid and pulmonary mass groups showed near-zero values with no flagged cases. Lung nodule detection identified at least one candidate in 78.2% of cases, totalling 292 candidates across the cohort (mean 3.36 ± 9.46 per case, mean diameter 7.68 ± 3.76 mm, 59.6% solid); the pneumothorax group showed the highest detection rate (100%) and mean count (8.15 ± 18.23 per case), with a high standard deviation suggesting an elevated false-positive rate, potentially driven by the heterogeneous density interfaces created by the partially collapsed lung parenchyma in pneumothorax cases.

Dataset 2: COVID-19 CT dataset (N = 20) — [placeholder]

The COVID-19 CT dataset is a publicly available collection of 20 thoracic CT scans from patients diagnosed with COVID-19,²⁴ assembled from two distinct sources: 10 scans acquired from coronacases.org and 10 scans sourced from radiopaedia.org. All scans are provided in NIfTI format and include expert segmentations of lung parenchyma and COVID-19 infection regions. Importantly, the two subsets differ markedly in their acquisition characteristics: the coronacases scans are isotropic volumetric acquisitions, whereas the radiopaedia scans were originally crawled as stacked axial JPEG/PNG images and subsequently converted to NIfTI, resulting in non-isotropic volumes with degraded inter-slice consistency. This heterogeneity, which reflects real-world variability in publicly available imaging data, provided an opportunity to assess pipeline robustness across input quality conditions.

Pipeline robustness and runtime

Of the 20 cases processed, 18 (90%) completed with full success across all three modules ( Table 3). The 2 partial failures (10%) were both from the Radiopaedia subset, in which the emphysema and nodule detection modules failed while TotalSegmentator completed successfully. Total batch processing time was 37.9 minutes, corresponding to a mean of approximately 1.9 minutes per case.

Across the 20 processed cases, results differed markedly between the two source subsets ( Table 4). TotalSegmentator flagged at least one pathological organ volume in 5.0% of cases overall (10.0% in the Coronacases subset, 0.0% in the Radiopaedia subset); all 20 cases had at least one undetected peripheral organ (mean 3.0 ± 1.05 per case), with the thyroid gland absent in 100% of cases across both subsets, consistent with the limited craniocaudal coverage of thoracic CT acquisitions. Heart volume could only be meaningfully computed in cases with valid segmentation output: mean was 478.86 ± 97.47 mL for Coronacases, while the Radiopaedia subset yielded largely invalid segmentations (mean 96.65 ± 86.75 mL), reflecting the non-isotropic volume geometry of these scans. Emphysema quantification was available only for the Coronacases subset (N = 10), yielding a mean LAA-950 of 0.43 ± 0.39% with no cases exceeding the pathological threshold; the module failed entirely for all Radiopaedia cases, again attributable to their non-standard acquisition geometry. Lung nodule detection similarly succeeded only in the Coronacases subset, with nodules detected in 90.0% of cases (mean 5.9 ± 4.87 per case, mean diameter 12.83 ± 5.52 mm, 94.9% solid, total 59 candidates); no nodules were detected in any Radiopaedia case, consistent with the module failure in this subset.