Viral Pneumonia Screening on Chest X-ray Images Using Confidence-Aware Anomaly Detection

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.

• The paper introduces a novel CAAD model that uses EfficientNet feature extraction and contrastive loss to detect viral pneumonia from chest X-rays.
• It demonstrates robust performance with an AUC of 87.57% on the X-VIRAL dataset and competitive results on unseen COVID-19 cases.
• The model’s confidence prediction enhances reliability, addressing challenges like dataset shift and class imbalance in clinical screenings.

Viral Pneumonia Screening on Chest X-rays Using Confidence-Aware Anomaly Detection

The paper presents a novel approach to viral pneumonia screening utilizing chest X-rays through a model known as Confidence-Aware Anomaly Detection (CAAD). This work reframes the task of distinguishing viral pneumonia from non-viral pneumonia and healthy controls into a one-class classification-based anomaly detection problem, addressing challenges such as dataset shift due to novel viruses and class imbalance between positive and negative cases.

Methodology

The CAAD model comprises several components: a shared feature extractor using EfficientNet, an anomaly detection module, and a confidence prediction module. The anomaly detection module assigns scores based on the likelihood of an X-ray being anomalous (indicative of viral pneumonia), while the confidence prediction module estimates the reliability of these anomaly scores.

To optimize the model, a contrastive loss function is implemented, leveraging anomaly data to reinforce the one-class model. The confidence prediction network then evaluates potential failures in anomaly detection by predicting the confidence level of anomaly scores.

Results

On the clinical X-VIRAL dataset containing 5,977 viral pneumonia and 37,393 non-viral or healthy cases, the CAAD model demonstrated superior performance relative to binary classification models. It achieved an Area Under the Curve (AUC) of 87.57% for identifying viral pneumonia. Furthermore, when tested on the X-COVID dataset without any prior exposure to COVID-19 instances, it achieved an AUC of 83.61% and a sensitivity of 71.70%. These outcomes are comparable to those observed in radiologists, underscoring the model's viability in real-world applications.

Implications

The CAAD model's reliance on anomaly detection rather than explicit classification addresses significant issues such as high intra-class variance and dataset shift, proving effective even under class-imbalance conditions. This approach enhances sensitivity, crucial for clinical settings where false negatives can have severe consequences. By predicting the confidence of anomaly detection, the model also reduces erroneous predictions, offering a layer of reliability essential for medical application.

Future Directions

The applicability of CAAD in large-scale, real-time medical screenings suggests significant potential for epidemic prevention and control. Future work could focus on refining the confidence prediction mechanism to reduce both false negatives and positives further. Moreover, integrating additional diagnostic information may enable assessments of pneumonia severity, facilitating early interventions for severe cases.

This research contributes a practical, adaptable framework for anomaly detection in medical imaging, particularly vital in contexts where rapid screening is required, and novel pathogens emerge. The findings suggest possible developments in AI models that are robust against evolving medical challenges.