AI Revolutionizes Otitis Media Diagnosis: A Game-Changer for Primary Care

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Sep 11 2023

In an article recently published in the Journal of Clinical Medicine, researchers performed a systematic review of artificial intelligence (AI) techniques used in automatic otitis media diagnostic studies based on medical images.

Background

Otitis media is a common disorder in children that leads to multiple symptoms, such as ear pain, sleep disturbances, and fever. Inaccurate diagnosis of otitis media can result in disease exacerbation, antibiotic overuse, unnecessary surgeries, cognitive development disorders, and hearing loss, which increases the importance of accurate diagnosis to ensure effective treatment and prevent greater complications.

However, accurate diagnosis of acute middle ear infections using the existing diagnostic techniques can be challenging. For instance, the narrow external ducts and the presence of earwax in infants can hinder accurate otitis media diagnosis using an ear endoscope. Moreover, the diagnosis accuracy can also be low due to unfamiliarity and systematic training with pneumatic ear endoscopy.

Different approaches, such as specialized medical student training programs, implementation of acoustic admittance and absorbance measurements, impedance measuring hearing aid integration, and development of novel otoscopic techniques and approaches, have been investigated to overcome these challenges.

However, these approaches did not increase the diagnostic success rates among otolaryngologists and pediatricians beyond 70% in primary care settings. Although medical image processing can play a crucial role in the exploration and analysis of medical data, the inherent complexity of medical images creates challenges in accurately evaluating and representing these images using conventional approaches.

Research has shown that AI can effectively analyze complex medical images. In otolaryngology, diagnostic accuracy varies based on the area of specialization and training of a physician due to the dependence on visual mechanisms and endoscopic imaging.

Deep learning (DL) algorithms can be integrated into oto-endoscopic imaging for image classification to increase diagnostic accuracy. However, DL-based automatic diagnostic systems have not been utilized for otitis media diagnosis in real-world clinical settings owing to the uncertainties related to DL.

AI in image-based otitis media diagnosis

In this paper, the authors systematically reviewed medical imaging-based automatic otitis media diagnostic studies that employed AI techniques to evaluate the feasibility of using AI to improve otitis media diagnosis accuracy. All relevant AI-assisted otitis media diagnosis studies were obtained from five databases, including IEEE Xplore, Embase, Medline, PubMed, and Google Scholar. Materials from reviewed studies and existing survey studies were compiled.

Literature searches on the databases were performed using a combination of AI-related and otitis-media-related keywords, including ‘eardrum’, ‘tympanic membrane disease’, ‘otitis media’, ‘otorhinolaryngology’, ‘ear abnormalities’, ‘middle ear’, ‘ear pathology’, ‘unsupervised learning’, ‘supervised learning’, ‘classification’, ‘segmentation’, ‘neural networks’, ‘convolutional neural networks’, ‘diagnose’, ‘computer diagnostics’, ‘automation’, ‘deep learning’, ‘machine learning’, and ‘artificial intelligence’.

Publish or Perish version 8 software was utilized to collect the literature from Google Scholar, while EndNote version X9 was used to eliminate duplicates from the obtained literature. Overall, 32 medical imaging studies were included in this review, with 26 studies using the classification technique, three employing the segmentation technique, and three utilizing both classification and segmentation techniques.

Computed tomography (CT), smartphone-based low-cost sword mirror, and oto-endoscopic images were used for classification studies, with the number of classification labels ranging from two to 14. Otoscopic/endoscopic images and only otoscopic images were used in studies using segmentation and segmentation and classification techniques, respectively.

Significance of the study

In classification studies, the random forest classifier in MATrix LABoratory (MATLAB) used for classifying optical coherence tomography (OCT) images displayed the highest otitis media diagnosis accuracy of 99.16%.

Other AI techniques that demonstrated a high otitis media diagnosis accuracy of above 90% in classification studies include multitasking joint sparse representation-based classification (MTJSRC), Xception, DenseNet169/1615/161, InceptionV3 and ResNet101, ResNet18+Shuffle, neural networks, random forest, and Alexnet, Googlenet, and Densenet201+ support vector machine (SVM). The single-shot multibox detector (SSD) showed the lowest accuracy of 48.7% among all 26 classification studies. Overall, the average otitis media diagnosis accuracy of classification studies was 86%.

In segmentation studies, both Mask R-convolutional neural networks (CNN) and EfficientNet-Attention-Residual U-Net (EAR-UNet) displayed a high diagnosis accuracy of 92.9% and 95.8% on endoscopic and otoscopic images, respectively. Classification with segmentation studies demonstrated 90.8% average otitis media diagnosis accuracy.

The model based on segmentation using Hough transform and classification using SVM showed the highest accuracy of 93.9% on otoscopic images, while the model based on segmentation using active contour models and classification using adaptive boosting (Adaboost) displayed the lowest accuracy of 88.06%. The average otitis media diagnosis accuracy of all studies reviewed in this paper was 86.5%, significantly higher than the 70% diagnostic accuracy of otolaryngologists and pediatricians in primary care settings.

To summarize, the findings of this review demonstrated the effectiveness of AI technologies in improving otitis media diagnosis accuracy, specifically in primary care settings and telemedicine.