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

TL;DR

This paper proposes cOOpD, a novel approach that reframes Chronic Obstructive Pulmonary Disease (COPD) classification on CT scans as an anomaly detection problem. By learning representations of healthy lung tissue and identifying deviations as anomalies, cOOpD achieves state-of-the-art performance and offers more interpretable results than traditional supervised methods, particularly for early-stage disease detection.

What Problem Does This Paper Address?

Diagnosing heterogeneous diseases like COPD is challenging due to their varied presentations and subtle appearances on medical imaging. COPD, in particular, is often underdiagnosed despite being a major cause of death. Traditional supervised machine learning models struggle with such diseases because they require extensive labeled data that accurately represents the full spectrum of disease manifestations. This is especially difficult for rare or early-stage cases, where the "diseased" class might be underrepresented or have subtle differences from healthy tissue. The authors propose to overcome this by treating COPD as an "anomaly" – deviations from what is considered normal lung tissue – rather than trying to learn all possible disease patterns directly.

Methods

The core of the cOOpD method involves two main stages:

  1. Self-Supervised Contrastive Representation Learning: The authors first train a deep learning model (specifically, a 3D ResNet encoder) using a self-supervised contrastive learning approach. This method learns to create meaningful "representations" (numerical summaries) of small 3D patches of lung tissue from CT scans. Crucially, this is done without explicit labels for disease. The contrastive learning objective encourages the model to group similar-looking lung tissue representations together and push dissimilar ones apart, effectively learning the intrinsic characteristics of lung tissue. This stage is designed to capture subtle features that might be indicative of disease or health.
  2. Anomaly Detection with a Generative Model: Once the encoder is trained, its learned representations are used for anomaly detection. A generative model (either a Gaussian Mixture Model or a Normalizing Flow) is trained only on the representations of healthy lung tissue patches. During inference, when a new lung patch is fed through the encoder, its representation is compared to the learned distribution of healthy representations. Patches whose representations fall outside this "normal" distribution are flagged as anomalous, indicating potential COPD. Patient-level scores are then derived by aggregating these patch-level anomaly scores.

Datasets

The study utilizes two large, publicly available, multi-center datasets for training and evaluation:

  • COPDGene: This dataset was used for training and internal validation. It was split into training (50%), validation (25%), and testing (25%) sets at the patient level.
  • COSYCONET: This dataset served as an independent external test set to evaluate the generalizability of the proposed method.

The authors also describe experiments with simulated real-world prevalence settings, where the COPDGene training set was undersampled to mimic lower disease prevalence.

Key Insights and Ideas

  • Reframing as Anomaly Detection is Powerful: The most significant insight is that treating COPD classification as an anomaly detection problem, rather than a direct supervised classification task, yields superior results. This approach is particularly effective for diseases with heterogeneous and subtle manifestations.
  • Latent Space is Key: Learning representations in a compressed "latent space" (the output of the contrastive encoder) is more effective than working directly with raw voxel data. This latent space captures essential features while reducing dimensionality and complexity.
  • Unsupervised Pre-training for Supervised Tasks: The self-supervised contrastive learning step is crucial. It allows the model to learn rich representations from unlabeled data, which then benefit the downstream anomaly detection task.
  • Interpretability Matters: The anomaly scores generated by cOOpD are not just predictive but also interpretable. The method can produce spatial anomaly maps, highlighting where in the lung the anomalies are detected, and patient-level scores that correlate with disease severity (GOLD stage), aiding in risk assessment and early detection.

Key Results

  • State-of-the-Art Performance: cOOpD outperformed all evaluated supervised state-of-the-art methods on both the internal (COPDGene) and external (COSYCONET) test sets.
  • Significant AUROC Improvement: cOOpD achieved an 8.2% and 7.7% increase in AUROC compared to the previous best supervised methods on COPDGene and COSYCONET, respectively.
  • Robustness to Prevalence: In simulated real-world prevalence scenarios (5%, 10.3%, 15% diseased class), cOOpD maintained its performance advantage over a strong supervised baseline (ReContrastive), which showed a significant performance drop.
  • Early Stage Detection Value: The anomaly scores showed a clear correlation with disease severity (GOLD stage), suggesting their utility in identifying individuals in the early stages of progression.

Limitations and Future Work

  • External Test Set Performance Drop: While cOOpD performed well on the external COSYCONET dataset, some supervised methods showed a larger performance drop. The authors hypothesize this is due to early signs of disease in the "healthy" class of COSYCONET, which might be better captured by the anomaly detection approach.
  • Future Directions: The authors suggest further validation on larger and more diverse datasets, longitudinal evaluation (tracking disease progression over time), and exploring the application of cOOpD to other heterogeneous diseases where labeled data is scarce but healthy data is abundant. They also note that the choice of generative model and aggregation strategy can influence performance.

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