AutoEncoder Convolutional Neural Network for Pneumonia Detection

TL;DR

This paper introduces an Autoencoder CNN approach for detecting pneumonia in paediatric chest x-rays, leveraging anomaly detection through histogram reconstruction error analysis to improve diagnostic accuracy.

Contribution

It presents a novel application of Autoencoder CNNs for pneumonia detection, including a specific error threshold for anomaly identification in medical imaging.

Findings

Distinct error differences between testing and training data

Threshold of 0.0127 effectively detects pneumonia anomalies

Autoencoder CNNs show promising diagnostic potential

Abstract

This study presents an innovative approach utilising Autoencoder Convolutional Neural Networks (AECNNs) for pneumonia detection in paediatric chest x-rays. The research addresses the complexity of pneumonia, considering diverse causative agents, including bacteria, viruses, and aspiration. Autoencoder Convolutional Neural Networks are employed to enhance anomaly detection by revealing hidden patterns in the data. The evaluation process involves meticulous analysis of the histogram reconstruction error, leading to the establishment of a threshold for anomaly identification. The results demonstrate distinct differences in error magnitudes during testing and training periods, with a threshold providing a tangible criterion for anomaly detection. The study contributes valuable insights into the discriminative capability of Autoencoder Convolutional Neural Networks, with a threshold of 0.0127, in detecting pneumonia in paediatric chest x-rays, emphasising their potential for improving diagnostic precision.