CT-3DFlow : Leveraging 3D Normalizing Flows for Unsupervised Detection of Pathological Pulmonary CT scans

TL;DR

CT-3DFlow addresses unsupervised pathology detection in chest CT by learning the normal distribution of healthy data with a 3D normalizing-flow model trained on $48×48×48$ patches. It leverages a 3D GLOW-based architecture with $L=4$ blocks and $K=64$ flows per block to produce per-patch log-likelihoods that are spatially aggregated into a patient-level decision, enabling anomaly detection without labeled pathology. The key contributions are the 3D NF architecture, patch-based training on $500,000$ normal patches, and a thresholding scheme for final prediction, showing superior performance against multiple baselines on a chest CT dataset of 822 patients. On the test set, the method achieved AUROC $0.952$, F1 $0.940$, and ACC $0.924$, demonstrating strong unsupervised anomaly detection with potential clinical impact and avenues for generalization to other modalities.

Abstract

Unsupervised pathology detection can be implemented by training a model on healthy data only and measuring the deviation from the training set upon inference, for example with CNN-based feature extraction and one-class classifiers, or reconstruction-score-based methods such as AEs, GANs and Diffusion models. Normalizing Flows (NF) have the ability to directly learn the probability distribution of training examples through an invertible architecture. We leverage this property in a novel 3D NF-based model named CT-3DFlow, specifically tailored for patient-level pulmonary pathology detection in chest CT data. Our model is trained unsupervised on healthy 3D pulmonary CT patches, and detects deviations from its log-likelihood distribution as anomalies. We aggregate patches-level likelihood values from a patient's CT scan to provide a patient-level 'normal'/'abnormal' prediction. Out-of-distribution detection performance is evaluated using expert annotations on a separate chest CT test dataset, outperforming other state-of-the-art methods.

Figures (1)

• Figure 1: Overview of the CT-3DFlow anomaly detection framework. The model, whose architecture is shown in the middle, is trained on normal 48x48x48 CT sub-volumes. During inference, overlapping patches are processed and aggregated to generate a full CT Log P map, which is binarized, filtered, and thresholded for classification.