[Paper Review] Viral Pneumonia Screening on Chest X-ray Images Using Confidence-Aware Anomaly Detection

Jianpeng Zhang, Yutong Xie+9 more|arXiv (Cornell University)|Mar 27, 2020

COVID-19 diagnosis using AIMedicine53 references132 citations

TL;DR

This paper reframes viral pneumonia detection on chest X-rays as a one-class anomaly detection problem and introduces CAAD, which combines anomaly scoring with a confidence predictor to improve sensitivity for screening viral pneumonia, including unseen pathogens like COVID-19.

ABSTRACT

Cluster of viral pneumonia occurrences during a short period of time may be a harbinger of an outbreak or pandemic, like SARS, MERS, and recent COVID-19. Rapid and accurate detection of viral pneumonia using chest X-ray can be significantly useful in large-scale screening and epidemic prevention, particularly when other chest imaging modalities are less available. Viral pneumonia often have diverse causes and exhibit notably different visual appearances on X-ray images. The evolution of viruses and the emergence of novel mutated viruses further result in substantial dataset shift, which greatly limits the performance of classification approaches. In this paper, we formulate the task of differentiating viral pneumonia from non-viral pneumonia and healthy controls into an one-class classification-based anomaly detection problem, and thus propose the confidence-aware anomaly detection (CAAD) model, which consists of a shared feature extractor, an anomaly detection module, and a confidence prediction module. If the anomaly score produced by the anomaly detection module is large enough or the confidence score estimated by the confidence prediction module is small enough, we accept the input as an anomaly case (i.e., viral pneumonia). The major advantage of our approach over binary classification is that we avoid modeling individual viral pneumonia classes explicitly and treat all known viral pneumonia cases as anomalies to reinforce the one-class model. The proposed model outperforms binary classification models on the clinical X-VIRAL dataset that contains 5,977 viral pneumonia (no COVID-19) cases, 18,619 non-viral pneumonia cases, and 18,774 healthy controls.

Motivation & Objective

• Motivate rapid viral pneumonia screening on chest X-rays to aid outbreak detection and control.
• Address dataset shift and extreme class imbalance by avoiding explicit multi-class viral labeling.
• Propose a one-class anomaly detection framework augmented with a confidence predictor to boost sensitivity.

Proposed method

• Shared feature extractor (EfficientNet-B0) feeding an anomaly detection module that outputs an anomaly score via a 3x100-neuron MLP.
• Contrastive loss using a reference Gaussian score to separate anomalies from normals.
• A separate confidence prediction network estimates a per-image confidence, calibrated via a probabilistic anomaly measure.
• Inference uses thresholds on anomaly score T_ano and confidence T_conf to decide positive/negative cases.
• Joint end-to-end training, with pretraining on ImageNet, and three-stage training (anomaly detection, confidence prediction, joint fine-tuning).
• Key equations include the anomaly score phi(x; theta, alpha), the contrastive loss L_ano, and the confidence formulation prob and g for failure prediction.

Experimental results

Research questions

• RQ1 Can viral pneumonia be effectively detected as an anomaly rather than via binary viral vs non-viral classification?
• RQ2 Does coupling an explicit confidence prediction module improve screening sensitivity and identify potential failures?
• RQ3 How well does CAAD generalize to unseen viral pneumonia cases (e.g., COVID-19) without fine-tuning?
• RQ4 What are the trade-offs between anomaly score thresholds and confidence thresholds for clinical screening?],
• RQ5 key_findings':['CAAD achieves state-of-the-art AUC on the X-VIRAL dataset (87.57% with confidence prediction) and high sensitivity (up to 93.01%) under certain confidence thresholds.
• RQ6 Anomaly detection outperforms binary classification baselines on the X-VIRAL dataset in terms of AUC and sensitivity (e.g., EfficientNet-B0: AUC 87.42% vs. 87.18% for the best binary model).
• RQ7 Confidence learning enables better failure prediction: anomaly probability more effectively separates correct vs. erroneous predictions than prediction probability.
• RQ8 On unseen X-COVID data (COVID-19 cases vs. normals) CAAD achieves AUC 83.61% and sensitivity 71.70%, comparable to radiologists in literature.
• RQ9 The model maintains robust performance without explicit COVID-19 training data during initial training, indicating good generalization for outbreak screening.

Key findings

• AUCs around 87% for X-VIRAL with CAAD vs. ~86% for binary baselines.
• Sensitivity improvements up to ~93% at higher confidence thresholds, with some loss in specificity.
• Confidence-aware re-labeling boosts screening sensitivity, supporting use in large-scale epidemic screening.