Robust multi-backbone hybrid fusion for Chest X-Ray pneumonia detection

Abstract

Chest X-ray imaging is extensively used for the diagnosis of pneumonia; however, manual interpretation is timeconsuming and prone to inter-observer variability. To overcome the limitations of fully supervised approaches and to improve stability, this paper introduces a multi-backbone hybrid framework that integrates supervised classification with unsupervised anomaly detection for pneumonia identification. The proposed method combines a supervised CNN-based classifier with a PatchCore anomaly detection model utilizing multiple pretrained backbones, including ResNet, DenseNet, and EfficientNet. Classifier confidence, predictive uncertainty, and PatchCore anomaly scores are fused using a logistic regression model trained exclusively on validation data, while strict leakage-free evaluation is ensured by keeping the test set fully locked. Experimental results on a publicly available chest X-ray dataset demonstrate that the proposed framework consistently achieves fused ROC-AUC scores exceeding 0.97 with low variance across five independent random seeds, indicating strong robustness, reproducibility, and suitability for reliable pneumonia screening.