Deep learning for automated diagnosis and differentiation of otitis media on temporal bone CT.
Authors
Affiliations (7)
Affiliations (7)
- School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China.
- Institute of Intelligent Chinese Medicine, Chongqing University of Chinese Medicine, Chongqing, China.
- International Research Center for Complexity Sciences, Hangzhou International Innovation Institute, Beihang University, Hangzhou, China.
- Department of Otolaryngology, Head and Neck Surgery, Affiliated Hospital of Southwest Medical University, Luzhou, China.
- Department of General Surgery (Hepatopancreatobiliary Surgery), Affiliated Hospital of Southwest Medical University, Luzhou, China.
- Department of Otolaryngology, Head and Neck Surgery, Affiliated Hospital of Southwest Medical University, Luzhou, China. [email protected].
- Department of Otolaryngology, Head and Neck Surgery, Affiliated Hospital of Southwest Medical University, Luzhou, China. [email protected].
Abstract
The accurate diagnosis of otitis media (OM) has long been a challenge for clinicians (especially for less experienced clinicians) due to the variety of types and the complex anatomical structures of the middle ear. Although deep learning (DL) based on different examination methods (mostly otoscopy) has been applied to the diagnosis of single species OM in previous studies, DL using temporal bone computed tomography (TBCT) images to diagnose OM and simultaneously differentiate between chronic otitis media (COM) and otitis media with effusion (OME) has not been investigated in depth. This study aimed to develop and evaluate a DL framework for the automated diagnosis of OM and identifying OME and COM with or without cholesteatoma using TBCT images. Our team created a unique large dataset of 2011 TBCT images from 1200 patients who were diagnosed with OM, which was determined the regions of interest (ROI) for middle ear (ME) by experienced experts. Then, a DL model was trained to detect the MEs in TBCT images and determine the OM status with this dataset of pre-processed images. Five-fold cross-validation was utilized for training and selecting the models. Finally, we evaluated the model using 406 images and verified the effectiveness of model-assisted diagnosis for different levels of clinicians in a comparative study. In the detection of the ME, the DL model achieved a detection ratio of 98.53%. The model showed satisfying performance in the classification of normal middle ear (NME), OME, and COM with an accuracy of 0.9238. With the assistance of the DL, the diagnostic accuracies were significantly improved from 81.53% to 93.60% (junior clinician) and from 87.93% to 95.57% (senior clinician), respectively. The findings suggested that the DL model could accurately identify MEs in TBCT images and classify NME, OME, and COM with satisfying accuracy. DL could also effectively assist clinicians in TBCT interpretation for OM diagnosis.