Projects per year
Abstract
Middle ear infection, also called otitis media, is extremely common in children with around 80% having a case before school age. It is challenging even for trained specialists to diagnose otitis media, especially in the subclassifications acute otitis media and otitis media with effusion. Untreated otitis media can cause hearing loss, delays in language acquisition, poor school performance, and behavioral problems. At the same time, otitis media is the leading contributor to antibiotic prescriptions and medical costs in children. Historically, there has been a global tendency to over-prescribe antibiotics in cases where middle ear effusion is present, but it is not clear if there is an infection. This PhD project aims to address the challenges of diagnosing otitis media by developing deep learning methods for automatic diagnosis. The work is based on a clinical dataset consisting of otoscopy images of the eardrum and wideband tympanometry measurements, which are objective measurements of the acoustic function of the middle ear. The contributions of this thesis are manifold. Three classification models are presented; one for the analysis of otoscopy images, one for the analysis of wideband tympanometry measurements, and a final approach based on a combination of the two modalities. It is shown that it is possible to determine the diagnosis based on these two different types of patient data using a deep learning model. Next, a generative model was developed for the generation of new artificial data from both modalities in the dataset. Additionally, it was examined how to employ a generative model to eliminate domain shifts in a medical image dataset. Domain shifts can occur when, e.g., data is collected in different hospitals or using different equipment. Furthermore, the human inter-rater variability of the diagnosis of the cases in the dataset was investigated. Each case was additionally diagnosed by four Ear-Noseand-Throat specialists based on the otoscopy images and wideband tympanometry measurements from the patients. This allowed for the determination of the diagnostic difficulty of each of the cases. A deep learning-based method for automatic estimation of the diagnostic difficulty was then developed.
The methods presented in this thesis could potentially be used as a diagnostic tool to assist medical professionals in the assessment of the condition of the eardrum,
and thus improve the diagnosis of otitis media in the future.
The methods presented in this thesis could potentially be used as a diagnostic tool to assist medical professionals in the assessment of the condition of the eardrum,
and thus improve the diagnosis of otitis media in the future.
Original language | English |
---|
Publisher | Technical University of Denmark |
---|---|
Number of pages | 132 |
Publication status | Published - 2022 |
Fingerprint
Dive into the research topics of 'Deep learning methods for pediatric middle ear diagnostics'. Together they form a unique fingerprint.Projects
- 1 Finished
-
Deep learning methods for pediatric middle ear diagnostics
Sundgaard, J. V. (PhD Student), Reyes, M. (Examiner), Paulsen, R. R. (Main Supervisor), Christensen, A. N. (Supervisor), Laugesen, S. (Supervisor), Dyrby, T. B. (Examiner) & Østergaard, L. R. (Examiner)
01/05/2019 → 12/09/2022
Project: PhD