Automatic Detection of Pathologies in Conventional Scanner Data

Abstract

Objectives

The thesis focuses on chest CT: a widely used modality with many examinations and highly reproducible volumetric data. The goal is a reading assistant that detects and localizes thoracic anomalies on any 3D CT scan with little annotation, by progressively reducing the level of supervision, from supervised to unsupervised methods.

I. Supervised: pulmonary embolism

Pulmonary embolism (PE) is a frequent and serious condition in which a clot blocks the pulmonary arteries. Diagnosis relies on CT angiography, the RV/LV diameter ratio and the Qanadli obstruction score. We compared two levels of supervision: a 3D segmentation network trained on fine clot segmentations from 54 PE-positive patients, and a 2D CNN + RNN classifier trained on image-level labels from more than 7,000 patients.

• Segmentation beats labels: learning from a few precise clot segmentations outperformed classification trained on thousands of image-level labels.
• Severity, not just presence: the same pipeline estimates the Qanadli score and the RV/LV ratio automatically.
• First place at the 2022 JFR data challenge (Journées Francophones de Radiologie) organized by the French Society of Radiology.

Full write-up: Detection and severity quantification of pulmonary embolism.

Supervised learning works very well for a specific pathology, but an anomaly detection system must cover every pathology and anatomical condition, and no representative annotated database exists. The next two parts reduce the need for annotation.

II. Self- and weakly supervised: semiological signs

Instead of one detector per disease, a semiological approach targets the patterns shared by most lung diseases: emphysema and cysts, nodules, consolidation and interstitial patterns. A 3D patch classifier is pre-trained with SimCLR contrastive learning on unlabeled patches from 1,337 patients, then fine-tuned on 4,000 visually selected patches from 1,330 patients. At inference, overlapping patches produce per-sign probability maps that are aggregated into a patient-level prediction.

• Self-supervision helps: SimCLR pre-training gave a clear boost in patient-level classification over random and ImageNet initialization, on a 251-patient test set.
• Simultaneous localization: the maps localize several abnormal regions at once, consistent with the ground-truth annotations.
• 88.8% recall and 88.3% precision of localization in a qualitative analysis by an expert radiologist.

Full write-ups: Weakly-supervised pathology detection and localization and Tailored 3D CT contrastive pretraining.

III. Unsupervised: normalizing flows

Normalizing flows are generative networks that learn the exact probability distribution of the data by transforming it into a simpler one. Trained on masked lung images from normal patients only, the model assigns low likelihood to anything that departs from normal anatomy. Thresholds and filters are selected on a validation set, and slice- or patch-level scores are aggregated into a patient-level decision.

The 3D variant, CT-3DFlow, works on 3D patches to catch pathologies that are hard to see in 2D or 2.5D and to localize them in 3D.

• Beats other unsupervised families: our flows outperform one-class SVM, autoencoders, GANs, diffusion models and other normalizing flows.
• 3D over 2D: 3D models perform better than the 2D versions.
• Close to self-supervision: without any annotation, the flows are competitive with the SimCLR model of Part II.
• Localization without supervision: the likelihood maps highlight the abnormal regions.

Full write-up: CT-3DFlow.

Conclusion & perspectives

• Less annotation: exploring several levels of supervision reduced the dependence on annotations while keeping good performance.
• Self-supervision and semiology: a promising route to generic reading-assistance systems.
• Unsupervised methods: an interesting alternative for anomaly detection, outperforming existing state-of-the-art methods.
• Current limits: performance does not yet allow routine clinical use, but opens promising perspectives.

Next steps include multimodal methods that combine imaging modalities (MRI, ultrasound, PET), text reports and biomarkers, and spectral CT, which adds information on tissue composition by separating materials by their absorption at different energies.

Jury

• Pr. Isabelle Bloch, Sorbonne Université · Reviewer
• Pr. Alexis Jacquier, Aix-Marseille Université · Reviewer
• Pr. Adrian Basarab, Université Claude Bernard Lyon 1 · Examiner
• Dr. Constance de Margerie-Mellon, Université Paris Cité · Examiner
• Dr. Olivier Nempont, Philips Research France · Co-supervisor
• Dr. Alexandre Popoff, Philips Research France · Co-supervisor
• Pr. Loïc Boussel, UCBL, Hospices Civils de Lyon · Thesis director

Publications from the thesis

1. Detection and severity quantification of pulmonary embolism with 3D CT data using an automated deep learning-based artificial solution, Diagnostic and Interventional Imaging, 2024.
2. Weakly-supervised learning-based pathology detection and localization in 3D chest CT scans, Medical Physics, 2024.
3. BERT-based natural language processing analysis of French CT reports: application to the measurement of the positivity rate for pulmonary embolism, Research in Diagnostic and Interventional Imaging, 2023.
4. eCAPRI: automated TAVI risk assessment with imaging biomarkers, European Radiology, 2025.
5. CT-3DFlow: Leveraging 3D Normalizing Flows for Unsupervised Detection of Pathological Pulmonary CT scans, arXiv, 2024.
6. Tailored 3D CT contrastive pretraining to improve pulmonary pathology classification, IEEE ICSP, 2022.

BibTeX

@phdthesis{djahnine2024thesis,
  title  = {D{\'e}tection automatique d'anomalies sur examens
            tomodensitom{\'e}triques standards},
  author = {Djahnine, Aissam},
  school = {Universit{\'e} Claude Bernard Lyon 1},
  year   = {2024},
  number = {2024LYO10167},
  url    = {https://theses.hal.science/tel-05263492}
}