[Paper Review] Early Diagnosis of Pneumonia with Deep Learning

Can Jozef Saul, Deniz Yagmur Urey+1 more|arXiv (Cornell University)|Apr 1, 2019

COVID-19 diagnosis using AI12 references45 citations

TL;DR

The paper presents a deep learning pipeline that preprocesses chest X-ray images and compares CNN and ResNet architectures, achieving 78.73% accuracy in pneumonia detection, surpassing CheXNet.

ABSTRACT

Pneumonia has been one of the fatal diseases and has the potential to result in severe consequences within a short period of time, due to the flow of fluid in lungs, which leads to drowning. If not acted upon by drugs at the right time, pneumonia may result in death of individuals. Therefore, the early diagnosis is a key factor along the progress of the disease. This paper focuses on the biological progress of pneumonia and its detection by x-ray imaging, overviews the studies conducted on enhancing the level of diagnosis, and presents the methodology and results of an automation of xray images based on various parameters in order to detect the disease at very early stages. In this study we propose our deep learning architecture for the classification task, which is trained with modified images, through multiple steps of preprocessing. Our classification method uses convolutional neural networks and residual network architecture for classifying the images. Our findings yield an accuracy of 78.73%, surpassing the previously top scoring accuracy of 76.8%.

Motivation & Objective

• Motivate the need for early pneumonia diagnosis and reduce human error in radiology.
• Propose an automated X-ray image preprocessing pipeline to enhance feature visibility.
• Evaluate CNN and Residual Network architectures for pneumonia classification on a public dataset.

Proposed method

• Apply three image preprocessing techniques: increase in color space, increase in contrast, and artificial lightening of images.
• Use a CNN with three convolutional layers and max pooling followed by a neural network classifier.
• Explore a 9-layer Residual Network (ResNet) for improved classification performance.
• Train with dropout, batch normalization, and Adam optimizer using binary cross-entropy loss across 120 epochs.
• Compare against CheXNet and ablations to assess preprocessing impact on accuracy.

Experimental results

Research questions

• RQ1 Can preprocessing of X-ray images improve pneumonia classification performance?
• RQ2 How do CNN and ResNet architectures compare on this pneumonia detection task?
• RQ3 Does combining contrast enhancement with image lightening yield higher accuracy than other preprocessing combinations?
• RQ4 How does the proposed approach perform relative to CheXNet on the same dataset?

Key findings

• The best model achieves 78.73% accuracy, surpassing CheXNet's 76.80%.
• Expanded color scheme alone provided modest gains over the base CNN.
• Increased contrast improved performance over the base model.
• Lightening on increased contrast yielded the highest accuracy among CNN variants before applying ResNet.
• A 9-layer ResNet further improves accuracy to the top value reported (78.73%).
• The study reports an F-score of 45.79% for their approach versus 38.7% for some conventional methods in cited work.