AI-Supported Comprehensive Detection and Quantification of Biomarkers of Subclinical Widespread Diseases at Chest CT for Preventive Medicine

Viktoria Palm, Tobias Norajitra, Oyunbileg von Stackelberg, Claus Peter Heußel, Stephan Skornitzke, Oliver Weinheimer, T. Kopytova, André Klein, Sílvia D. Almeida, Michael Baumgartner, Dimitrios Bounias, Jonas Scherer, Klaus Kades, Hanno Gao, Paul F. Jäger, Marco Nolden, Elizabeth Tong, Kira Eckl, Johanna Nattenmüller, Tobias Nonnenmacher, Omar Naas, Julia Reuter, Arved Bischoff, Jonas Kroschke, Fabian Rengier, Kai Schlamp, Manuel Debic, Hans‐Ulrich Kauczor, Klaus Maier‐Hein, Mark O. WielpützHealthcare 2022 · DOI

TL;DR

This paper introduces "Comorbidome," a comprehensive AI system that analyzes chest CT scans to simultaneously detect and quantify multiple subclinical diseases (pulmonary, metabolic, cardiovascular, musculoskeletal). This holistic approach aims to provide a more complete picture of a patient's health, enabling earlier intervention and preventive medicine, moving beyond single-disease AI tools.

What Problem Does This Paper Address?

Current medical practice often treats diseases in isolation, and even AI tools in radiology tend to focus on a single condition (e.g., detecting a lung nodule or assessing emphysema). However, patients frequently present with multiple, interconnected, and often subclinical health issues that can significantly impact their overall prognosis. These "comorbidities" can be subtle and difficult for the human eye to consistently detect and quantify, especially when they are not the primary reason for the CT scan. The challenge is to move towards a more integrated diagnostic approach that can identify and assess these co-occurring conditions efficiently, paving the way for preventive medicine and better patient management.

Methods

The core of the Comorbidome project is a multidimensional AI workflow built on a decentralized Joint Imaging Platform (JIP) using the Kaapana toolkit. This platform orchestrates various deep learning algorithms, each specialized for detecting and quantifying specific comorbidities. Key aspects of their methodology include:

  • Multidisciplinary Team: A collaborative effort involving radiologists, computer scientists, physicists, and mathematicians.
  • Deep Learning Framework: Primarily utilizes the nnU-Net framework for its automated configuration, optimization, and training capabilities in medical image segmentation. For object detection, nnDetection is employed.
  • Synchronized Workflow: Algorithms run concurrently on the JIP, processing different aspects of the CT scan simultaneously.
  • Structured Reporting: Results are compiled into a structured report ("hAlco" report) that includes quantitative findings, reference values, recommendations, and annotated sample images.
  • Data Augmentation: Standard techniques like random rotations, scaling, elastic deformations, gamma correction, and mirroring are used to enhance model robustness.
  • Preprocessing: Includes clipping image intensities and z-score normalization.

Datasets

The study utilized a large, multicentric dataset of over 10,000 CT datasets collected retrospectively over five years (January 2016–December 2020) from seven different CT scanners at the University Hospital Heidelberg and Thoraxklinik Heidelberg. This heterogeneity in acquisition parameters was intended to ensure a robust training dataset. For specific tasks, publicly available datasets were also used for pre-training models.

Key Insights and Ideas

  • Holistic Imaging Diagnostics: The central idea is that a single chest CT scan contains a wealth of information beyond the primary indication, and AI can unlock this potential for comprehensive health assessment.
  • "Comorbidome" Concept: The project aims to create a "comorbidome" – a complete picture of a patient's comorbidities – rather than just focusing on individual diseases.
  • Preventive Medicine Focus: By detecting subclinical conditions, the tool can facilitate early interventions, lifestyle adjustments, and personalized preventive strategies.
  • Bridging the Gap: The system aims to bridge the gap between qualitative visual assessment and the quantitative data needed for precise risk stratification and management.
  • Scientific Exploration: The platform opens new avenues for research into previously unknown correlations between different diseases and their impact on health outcomes.

Key Results

The paper demonstrates the feasibility of the Comorbidome system through two example cases:

  • TAVI Patient: An automated analysis of a 75-year-old male patient undergoing assessment for Transcatheter Aortic Valve Implantation (TAVI) revealed abnormal visceral fat area (VFA), indicating a need for nutritional adjustment. High coronary and aortic calcium scores suggested a high risk for future cardiac events, recommending statin therapy. An incidental pulmonary nodule was detected and classified as Lung-RADS grade 2, requiring follow-up. Bone mineral density showed significantly reduced t-scores, necessitating further diagnostic workup.
  • COPD Patient: An automated evaluation of a 66-year-old male patient for COPD diagnosis identified moderate parenchymal lung changes, mild emphysema, and bronchial wall thickening. Intermediate hepatic steatosis was noted, along with a moderate coronary artery calcium score, again indicating a high risk for cardiac events and recommending statin therapy. Bone mineral density was significantly reduced, suggesting osteoporosis and requiring diagnostic workup.

The system generates structured reports that highlight findings in red (abnormal), yellow (intermediate), and green (normal), aiding intuitive interpretation.

Limitations and Future Work

  • Validation: The paper presents the concept and initial results, acknowledging that further detailed evaluation and validation of each algorithm are ongoing.
  • Clinical Integration: While the infrastructure is designed for clinical routine, the full integration and impact on patient care require further study.
  • Future Directions:
    • Integration of healthcare data from digital documentation.
  • Extension of the osteoporosis evaluation tool to include vertebra volume, fracture grading, and microarchitecture analysis.
  • Development of algorithms for identifying subtle imaging biomarkers of malignancy risk in pulmonary nodules.
  • Analysis of lung parenchyma for fibrotic changes.
  • Extraction and evaluation of radiomic features.
  • Potential fusion with other AI tools (e.g., for abdominal CT analysis) to create even more comprehensive systems.

Summary updated

Abstract

Automated image analysis plays an increasing role in radiology in detecting and quantifying image features outside of the perception of human eyes. Common AI-based approaches address a single medical problem, although patients often present with multiple interacting, frequently subclinical medical conditions. A holistic imaging diagnostics tool based on artificial intelligence (AI) has the potential of providing an overview of multi-system comorbidities within a single workflow. An interdisciplinary, multicentric team of medical experts and computer scientists designed a pipeline, comprising AI-based tools for the automated detection, quantification and characterization of the most common pulmonary, metabolic, cardiovascular and musculoskeletal comorbidities in chest computed tomography (CT). To provide a comprehensive evaluation of each patient, a multidimensional workflow was established with algorithms operating synchronously on a decentralized Joined Imaging Platform (JIP). The results of each patient are transferred to a dedicated database and summarized as a structured report with reference to available reference values and annotated sample images of detected pathologies. Hence, this tool allows for the comprehensive, large-scale analysis of imaging-biomarkers of comorbidities in chest CT, first in science and then in clinical routine. Moreover, this tool accommodates the quantitative analysis and classification of each pathology, providing integral diagnostic and prognostic value, and subsequently leading to improved preventive patient care and further possibilities for future studies.
BibTeX
@inproceedings{palm2022ai,
  author = {Viktoria Palm and Tobias Norajitra and Oyunbileg von Stackelberg and Claus Peter Heußel and Stephan Skornitzke and Oliver Weinheimer and T. Kopytova and André Klein and Sílvia D. Almeida and Michael Baumgartner and Dimitrios Bounias and Jonas Scherer and Klaus Kades and Hanno Gao and Paul F. Jäger and Marco Nolden and Elizabeth Tong and Kira Eckl and Johanna Nattenmüller and Tobias Nonnenmacher and Omar Naas and Julia Reuter and Arved Bischoff and Jonas Kroschke and Fabian Rengier and Kai Schlamp and Manuel Debic and Hans‐Ulrich Kauczor and Klaus Maier‐Hein and Mark O. Wielpütz},
  title = {{AI-Supported Comprehensive Detection and Quantification of Biomarkers of Subclinical Widespread Diseases at Chest CT for Preventive Medicine}},
  booktitle = {Healthcare},
  year = {2022},
  doi = {10.3390/healthcare10112166},
}
Figure