[Paper Review] Predicting Pneumonia and Region Detection from X-Ray Images using Deep Neural Network

Sheikh Md Hanif Hossain, S M Raju+1 more|arXiv (Cornell University)|Jan 19, 2021

COVID-19 diagnosis using AI4 citations

TL;DR

This paper proposes a deep learning approach using transfer learning with pre-trained ResNet-50 to detect pneumonia from chest X-ray images and localize infected lung regions. The model achieves 90.6% accuracy and employs class activation mapping for interpretability, with a user-friendly tool, PneuNet, developed for clinical accessibility.

ABSTRACT

Biomedical images are increasing drastically. Along the way, many machine learning algorithms have been proposed to predict and identify various kinds of diseases. One such disease is Pneumonia which is an infection caused by both bacteria and viruses through the inflammation of a person's lung air sacs. In this paper, an algorithm was proposed that receives x-ray images as input and verifies whether this patient is infected by Pneumonia as well as specific region of the lungs that the inflammation has occurred at. The algorithm is based on the transfer learning mechanism where pre-trained ResNet-50 (Convolutional Neural Network) was used followed by some custom layer for making the prediction. The model has achieved an accuracy of 90.6 percent which confirms that the model is effective and can be implemented for the detection of Pneumonia in patients. Furthermore, a class activation map is used for the detection of the infected region in the lungs. Also, PneuNet was developed so that users can access more easily and use the services.

Motivation & Objective

• To develop an automated system for detecting pneumonia in chest X-ray images using deep neural networks.
• To localize the specific regions in the lungs affected by pneumonia using explainable AI techniques.
• To improve diagnostic accuracy and clinical utility through transfer learning with pre-trained models.
• To create a user-accessible tool, PneuNet, for real-world deployment in healthcare settings.

Proposed method

• Transfer learning was applied using a pre-trained ResNet-50 convolutional neural network as the feature extractor.
• Custom dense and dropout layers were added after the feature extraction stage for binary classification of pneumonia.
• Class activation mapping (CAM) was used to visualize and localize regions of inflammation in the lungs.
• The model was trained and validated on a public chest X-ray dataset with balanced pneumonia and normal image classes.
• A web-based interface, PneuNet, was developed to allow clinicians and researchers to upload images and receive predictions.
• Data augmentation techniques were applied to improve model generalization and reduce overfitting.

Experimental results

Research questions

• RQ1 Can a deep learning model based on transfer learning accurately classify pneumonia from chest X-ray images?
• RQ2 Can the model localize the specific regions of lung inflammation using attention-based visualization techniques?
• RQ3 How does the performance of the proposed model compare to existing methods in terms of accuracy and interpretability?
• RQ4 Can a user-friendly interface be effectively developed to support clinical deployment of the model?

Key findings

• The proposed model achieved a classification accuracy of 90.6% on the test set, demonstrating strong predictive performance.
• Class activation mapping successfully highlighted regions of lung inflammation, enhancing model interpretability and clinical trust.
• The integration of transfer learning with ResNet-50 significantly improved training efficiency and model generalization.
• The developed PneuNet platform enables accessible and real-time prediction for medical professionals and researchers.
• The model demonstrated robustness to variations in image quality and anatomical positioning through data augmentation.