Author: Phong Phu Nguyen
Nguyen, Phong Phu, 2021 Developing a deep learning algorithm to improve diagnosis of otitis media, Flinders University, College of Science and Engineering
Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.
Background: Otitis Media is a common childhood ailment. In clinical practice, the diagnosis of Otitis Media includes the visual examination of the tympanic membrane from otoscope (otoscopy images) and measuring the movement patterns of the eardrum for different pressure (tympanometry data). However, the diagnosis requires extensive tool usage training and result interpretation. Misdiagnosis is a common problem for community-based clinicians in remote areas, which could cause delays in the treatment of the disease. The report proposes machine learning systems that could give predictions on different types of otitis media by using multiple sources of diagnosis data available.
Methods: Support Vector Machine, Multi-Layer Perceptron and Convolutional Neural Network models are used to build the system on two different datasets. The first dataset is a public one, with 454 images of three different categories. The second dataset is the Swimming Pool database coming from a report in regional South Australia for child ear diagnosis. Data is available in terms of otoscopy videos and various diagnosis labelling. Input processing and normalisation, principal component analysis, and grid search have been utilised for training the model with available data in both multi-label and binary classification problems. The tympanometry data has been combined with the prediction results from extracted otoscopy images to boost the overall performance of the algorithm.
Results: The best model accuracy achieved on 3-label problem on the public dataset is 83%. The highest accuracies for a 5-label classification problem and binary classification problem on Swimming Pool data are 64% and 78%, respectively. By combining the tympanometry data with the probability prediction of otoscopy-image-based model, the accuracy of the system has been increased from 78% to 82% in the binary classification.
Conclusion: Machine learning models could be used to build a system that could support the otitis media diagnosis to support effective triage, timely patient referrals and effective treatment of the illness. The report also suggests the possibility of combining different diagnosis data types to increase the overall predictive performance of the system.
Keywords: Otitis Media, Machine Learning, Deep Learning, Support Vector Machine, SVM, Multilayer Perceptron, MLP, Convolutional Neural Network, CNN, Otoscopy Video, Otoscopy Image, Tympanometry Data, Combining Diagnosis Data
Subject: Computer Science thesis
Thesis type: Masters
Completed: 2021
School: College of Science and Engineering
Supervisor: Dr Trent Lewis