P416 Efficient and accurate diagnosis of otomycosis using an ensemble deep learning model

Abstract

Poster session 3, September 23, 2022, 12:30 PM - 1:30 PMObjectivesOtomycosis accounts for ˃15% of the cases with external otitis worldwide. And otomycosis is more frequently observed in humid regions and people enjoying the culture of ear cleaning in China. Aspergillus and Candida are the major pathogens that could cause long-term infection. Early endoscopic and microbiological examinations are important for appropriate medical treatment to otomycosis. However, accurate diagnosis always needs experts such as otologist and microbiologist. Deep learning model is a novel efficient method to provide quick diagnosis which is an automatically diagnostic program using a large database of images acquired in the cliniC. This paper puts forward a mechanic learning model to address the diagnosis of otomycosis caused by Aspergillus and Candida accurately and quickly.MethodsWe proposed a computer-aided decision system that is based on a deep learning model consisting of two subsystems, a java-based web application, and picture classification. The web application subsystem mainly provides a user-friendly page for collecting consulted pictures as well as displaying the calculation results. The picture classification subsystem mainly uses trained neural network models for end-to-end data inference. The end user only needs to upload a few pictures of the ear endoscope, and the system will return the classification results to the user in the form of category probability value.In order to accurately diagnose otomycosis, we generally kept endoscopic images and took the secretion for fungal culture for further identification. Positive fluorescence fungal staining, culture, and further DNA sequencing were taken to confirm the pathogens, Aspergillus or Candida sp. In addition, impacted cerumen, external otitis, and normal external auditory canal endoscopic images are retained for reference. We merged these four types of images into an endoscopic images gallery.ResultsIn order to achieve better accuracy and generalization ability after model training, we selected 2750 samples from nearly 4000 ear endoscopic images as training samples and 454 as validation samples. On the selection of deep neural network models, we tested the resnet, senet, and efficientnet neural network models with different numbers of layers. Considering the accuracy and operation speed, we finally chose the efficientnet-b6 model and output the probability values of the four categories of otomycosis, external otitis, impacted cerumen, and normal cases. After multiple iterative sample training, the overall validation sample accuracy reached 94.71%, and the average cross-validation accuracy of the 4 classifications reached 94.3%.ConclusionThe results suggest that the system can be used as a reference for general practitioners to make better decisions in the diagnosis of otomycosis.