Detecting Non-CXR Inputs via Validation-Calibrated Thresholds in a Three-Level Convolutional Autoencoder for Chest Radiographs

Abstract

Ensuring that images submitted to web-based diagnostic systems are genuine chest radiographs (CXR) is essential to prevent errors in automated disease detection workflows. This work proposes a three-level convolutional autoencoder trained exclusively on CXR images to identify non-CXR inputs through reconstruction-error analysis. A threshold calibrated on the 95th percentile of the validation MSE distribution (0.000960) was used to determine class membership. The model achieved balanced accuracies above 0.89 across multiple test domains, including medical (BUSI, Dental OPG, MildDemented) and non-medical datasets (ImageNet, Baggage X-ray). The results confirm the model’s capability to act as a lightweight, unsupervised gate that filters off-target images before diagnostic inference. This approach supports data integrity, reliability, and operational safety in CXR-based disease detection systems.