Abstract: Reformulating COPD Classification on Chest CT Scans as Anomaly Detection using Contrastive Representations - cOOpD

Silvia Dias Almeida, Carsten T. Lüth, Tobias Norajitra, Tassilo Wald, Marco Nolden, Paul F. Jäger, Claus Peter Heussel, Jürgen Biederer, Oliver Weinheimer, Klaus H. Maier-Hein

Introduction

Chronic Obstructive Pulmonary Disease (COPD) is a leading cause of death worldwide, yet it remains significantly underdiagnosed. Detecting COPD via Computed Tomography (CT) is inherently difficult because the disease presents in a sparse, diffuse, and highly heterogeneous manner. Traditional machine learning approaches typically treat this as a supervised binary classification task (COPD vs. Healthy). However, because the visual "signature" of COPD is so varied across patients, supervised models often struggle to generalize to the full spectrum of the disease.

This research introduces cOOpD, a framework that reformulates COPD detection as an anomaly detection task. Instead of trying to learn every possible appearance of the disease, the authors propose learning a robust representation of what healthy lung tissue looks like and flagging deviations from this norm as anomalies. This approach is particularly suited for heterogeneous diseases where the "abnormal" class is difficult to define comprehensively.

Methodology

The cOOpD framework utilizes a multi-stage pipeline designed to move away from "black box" supervised classification toward a more flexible and interpretable "one-class" style of detection:

  1. Self-Supervised Representation Learning: The model uses a contrastive pretext model to learn feature representations of unlabeled lung regions. This allows the system to capture underlying tissue characteristics and anatomical structures without requiring extensive manual expert annotations, which are often expensive and time-consuming to obtain.
  2. Generative Modeling of Normalcy: A generative model is trained specifically on the representations of healthy lung regions. This stage establishes a statistical distribution or "manifold" of what constitutes normal lung tissue in the learned feature space.
  3. Anomaly Detection: Pathological regions are identified as Out-of-Distribution (OOD) samples. When the model encounters lung tissue that significantly differs from the learned healthy distribution, it assigns a high anomaly score to that region.
  4. Aggregation: These region-level OOD scores are aggregated across the entire lung volume to produce a final patient-level score for diagnosis. This allows the model to provide both a global assessment and localized findings.

Key Insights and Interpretability

By focusing on the consistency of healthy tissue, the model becomes more sensitive to the "chaos" and structural changes associated with diseased tissue. This approach offers several critical advantages for clinical use:

  • Embracing Heterogeneity: Rather than viewing the varied appearance of COPD as a hurdle, cOOpD treats it as a defining characteristic of its deviation from normalcy. This makes the model more robust to rare or unusual disease presentations.
  • Spatial Interpretability: Unlike standard classifiers that output a single probability score, cOOpD generates spatial anomaly maps. These maps highlight exactly where in the lungs the abnormalities are located, providing clinicians with actionable visual evidence that can be correlated with physiological symptoms.
  • Early Detection Value: The authors found that their method is particularly effective at identifying individuals in the early stages of disease progression. This is a critical window for clinical intervention where supervised models often fail because the visual signs of disease are subtle.
  • Prevalence Robustness: The model performs well across different disease prevalence settings. This suggests it is better suited for real-world clinical deployment, where the ratio of healthy to diseased patients can fluctuate significantly, unlike the balanced datasets often used in laboratory settings.

Results and Performance

The study evaluated cOOpD on two public datasets, testing its robustness in artificially designed real-world prevalence settings. The model achieved a significant performance boost over previous supervised state-of-the-art methods, with an increase in AUROC (Area Under the Receiver Operating Characteristic curve) of 8.2% and 7.7% on the respective datasets.

Qualitatively, the model produced highly interpretable spatial maps that align well with known pathological regions. This success in both quantitative metrics and qualitative visualization underscores the potential of anomaly detection as a superior paradigm for complex medical imaging tasks.

Conclusion

By reframing COPD classification as an anomaly detection task using self-supervised contrastive representations, the cOOpD framework provides a more robust, interpretable, and clinically relevant approach to diagnosing heterogeneous lung diseases. Its ability to detect early-stage disease and provide localized visual evidence marks a significant step forward in the automated analysis of chest CT scans, offering a promising path toward earlier intervention and better patient outcomes.

Summary updated

Abstract

Classification of heterogeneous diseases is challenging due to their complexity, variability of symptoms and imaging findings. Chronic obstructive pulmonary disease (COPD) is a prime example, being underdiagnosed despite being the third leading cause of death. Its sparse, diffuse and heterogeneous appearance on computed tomography challenges supervised binary classification. We reformulate COPD binary classification as an anomaly detection task, proposing cOOpD: heterogeneous pathological regions are detected as out-of-distribution (OOD) from normal homogeneous lung regions. To this end, we learn representations of unlabeled lung regions employing a self-supervised contrastive pretext model, potentially capturing specific characteristics of diseased and healthy unlabeled regions.Agenerative model then learns the distribution of healthy representations and identifies abnormalities (stemming from COPD) as deviations. Patientlevel scores are obtained by aggregating region OOD scores. We show that cOOpD achieves the best performance on two public datasets, with an increase of 8.2% and 7.7% in terms of AUROC compared to the previous supervised state-of-the-art. Additionally, cOOpD yields well-interpretable spatial anomaly maps and patient-level scores which we show to be of additional value in identifying individuals in the early stage of progression. Experiments in artificially designed real-world prevalence settings further support that anomaly detection is a powerful way of tackling COPD classification [1].
BibTeX
@inproceedings{almeida2024abstract,
  author = {Silvia Dias Almeida and Carsten T. Lüth and Tobias Norajitra and Tassilo Wald and Marco Nolden and Paul F. Jäger and Claus Peter Heussel and Jürgen Biederer and Oliver Weinheimer and Klaus H. Maier-Hein},
  title = {{Abstract: Reformulating COPD Classification on Chest CT Scans as Anomaly Detection using Contrastive Representations - cOOpD}},
  booktitle = {Bildverarbeitung für die Medizin},
  year = {2024},
  doi = {10.1007/978-3-658-44037-4\_34},
}
Figure