| dc.description.abstract | Abstract
Mastoid abnormalities show different types of ear diseases, time consuming and low accuracy in diagnosing some small abnormal mastoids, requires a new approach to diagnose these abnormalities and reduce human error. The main disadvantages of manual Mastoid CT scan are time-consuming and error-prone. This study presents for the first time a model for diagnosing mastoid abnormalities based on deep learning approaches using a large database of high-precision clinical CT images.
Materials and methods
In this study, the data set used included 24,800 CT scan slides (right and left mastoid) from 152 patients referred to Golgasht Imaging Center in Tabriz at the request of an ENT specialist, which included mastoid air cells from the highest to the lowest part of the ear cavity. includes.
This project was done in two general parts:
1) Classification of healthy and unhealthy mastoid air cells (2 classes)
2) Classification of healthy and 4 unhealthy types of mastoid air cells (5 classes)
In the first part, a powerful method based on convolution layers and deep neural network for classifying normal and abnormal mastoid air cells is introduced, which has better results than well-known networks such as ResNet and AlexNet.
In the second part, a wide range of different architectures with different optimization methods were used to classify natural Mastoid CT scan images and four different types of abnormalities, and the best type of architecture and optimization function with the highest percentage in terms of accuracy was selected.
Results
In the first part, the proposed fully automatic classification and detection method provides a more efficient and faster result compared to the ENT specialist manual classification. In the proposed algorithm, the evaluation criteria of accuracy, score 1, accuracy and recovery were 98.10%, 98.05%, 98.32%, 97.89% (average in five-fold cross-validation, respectively), which was higher than the results obtained from the resonant network and the problem of overfitting Has also been fixed.
In the second part, the classification of 5 classes with Xception architecture and Adamax cost function was obtained with an average accuracy of 87.70%, which is the highest value compared to other architectures and functions.
Conclusion
Manual analysis of CT scan of the ear canal is often time consuming and error prone due to various studies of variation between the operator or within the operator. The proposed method can be used for automatic analysis of the middle ear cavity to classify mastoid abnormalities, which are significantly faster than manual diagnosis, have very low error rates, and the results obtained in the two-class classification were more accurate than the ResNet and AlexNet networks. In the five-class classification, the Xception architecture with the Adamax cost function also performed best. | en_US |
