Presentation Title

Diagnosis of Chest Diseases using Neural Networks and Deep Learning

Faculty Mentor

Dr. Paul Wilkinson

Start Date

17-11-2018 9:00 AM

End Date

17-11-2018 9:15 AM

Location

C162

Session

Oral 1

Type of Presentation

Oral Talk

Subject Area

engineering_computer_science

Abstract

The prominence of neural networks in areas such as marketing and automation has led its application in other industries. Among these sectors being introduced to deep learning is healthcare. The healthcare environment provides an abundance of data, however, it lacks effective analysis utilities to discover hidden relationships and trends. This research explores the techniques of deep learning to classify chest X-rays, provided by the National Institutes for Health, with diseases, more specifically, tuberculosis.

Artificial intelligence provides computers with the ability to determine hidden patterns invisible to the human eye. A neural network, structured similarly to the human brain, is composed of dense layers of neurons. This structure allows for a model to be trained and optimized through back propagation as inputs move through each layer of the network. Because a linear model or clustering algorithms would present major inaccuracies, activation functions such as the rectified linear units can be implemented in order to provide efficient regression and classification. With the introduction of Google’s Tensorflow library, a convolutional neural network was trained throughout this research with over 400 images to classify X-rays diagnosed with tuberculosis. The minimization of parameters from a convolutional neural network significantly increases efficiency for image classification.

Utilizing a separate dataset of 35 normal and 35 abnormal X-rays, the neural model was evaluated for both precision and accuracy. Comparing every RGB value, each X-ray yielded explicit certainty amounts which were used to measure the precisions of the normal and abnormal sets. Additionally, all measurements over 50% certainty were considered successful, resulting in a 77% accuracy rate after 4000 epochs of training. With the addition of GPU’s, larger uniform datasets, and a powerful computer, neural networks as a means of diagnosing diseases become more practical as they are able to analyze complexities effectively.

Summary of research results to be presented

This research mainly revolved around how deep learning could be utilized to automate the medical diagnosis process. After creating a convolutional neural network using Google’s Tensorflow, a classification model was trained with multiple datasets containing chest X-rays with tuberculosis and normal X-rays. Then, 35 normal and 35 abnormal X-rays were tested.

The first neural network was trained with about 30 normal and abnormal X-rays. The model trained for 500 epochs diagnosed 77% of the normal X-rays and 24% of the abnormal X-rays accurately. The model trained for 4000 epochs diagnosed 80% of the normal X-rays and 34% of the abnormal X-rays accurately.

The model was trained on only about 30 images each, which may present large amounts of error since new disparities between pixels may arise. It would most likely be more accurate with at least 100 images to train. An additional error may have occurred due to training X-rays between different hospitals since both hospitals formatted their data differently.

Using about 200 images of both normal and abnormal X-rays from only one hospital, and testing the same dataset after 4000 epochs results in 77% accuracy for both sets of X-rays. This trial was significantly more successful due to the limitations of errors as described above.

There are many technological difficulties to overcome before this process could become normalized. Once machine learning algorithms become more advanced, accounting for less loss and greater optimization, the process of automating the medical diagnosis process will become more practical.

This document is currently not available here.

Share

COinS
 
Nov 17th, 9:00 AM Nov 17th, 9:15 AM

Diagnosis of Chest Diseases using Neural Networks and Deep Learning

C162

The prominence of neural networks in areas such as marketing and automation has led its application in other industries. Among these sectors being introduced to deep learning is healthcare. The healthcare environment provides an abundance of data, however, it lacks effective analysis utilities to discover hidden relationships and trends. This research explores the techniques of deep learning to classify chest X-rays, provided by the National Institutes for Health, with diseases, more specifically, tuberculosis.

Artificial intelligence provides computers with the ability to determine hidden patterns invisible to the human eye. A neural network, structured similarly to the human brain, is composed of dense layers of neurons. This structure allows for a model to be trained and optimized through back propagation as inputs move through each layer of the network. Because a linear model or clustering algorithms would present major inaccuracies, activation functions such as the rectified linear units can be implemented in order to provide efficient regression and classification. With the introduction of Google’s Tensorflow library, a convolutional neural network was trained throughout this research with over 400 images to classify X-rays diagnosed with tuberculosis. The minimization of parameters from a convolutional neural network significantly increases efficiency for image classification.

Utilizing a separate dataset of 35 normal and 35 abnormal X-rays, the neural model was evaluated for both precision and accuracy. Comparing every RGB value, each X-ray yielded explicit certainty amounts which were used to measure the precisions of the normal and abnormal sets. Additionally, all measurements over 50% certainty were considered successful, resulting in a 77% accuracy rate after 4000 epochs of training. With the addition of GPU’s, larger uniform datasets, and a powerful computer, neural networks as a means of diagnosing diseases become more practical as they are able to analyze complexities effectively.