|
|
 |
| |
|
|
|
Automatic Sarcoidosis Stage Classification Based on Gray Level Co-occurrence Matrix Features |
|
|
|
PP: 197-208 |
|
|
|
doi:10.18576/amis/190117
|
|
|
|
Author(s) |
|
|
|
Mohanad A. Deif,
Hani Attar,
Mohamed A. Hafez,
Waleed Alomoush,
Hussein Al-Faiz,
|
|
|
|
Abstract |
|
|
| The correct diagnosis and staging of this complex inflammatory disease that majorly affects the lungs, such as Sarcoidosis, are very important. Therefore, this research focuses on differentiating the four stages of Sarcoidosis using chest X-ray images with the application of three machine learning classifiers: K-Nearest Neighbours, Support Vector Machine, and Artificial Neural Network. GLCM was used for feature extraction after segmentation using Otsus method, and K-means clustering was used to enhance feature reliability and accuracy. This shows that, in the case of using Otsus, the distinction of the stages of Sarcoidosis is better since our analysis showed higher average Jaccard Indexes than K- means and no segmentation. Performance evaluation for the classifiers was done using several metrics, including accuracy, precision, recall, and the F1 score. It was found that the results, generally across most of the modelled stages, performed well using both KNN and SVM. However, for Stage 2, KNN produced the best result, producing an accuracy of 97.83%, while SVM immediately did so for Stage 4, creating an accuracy of 97.1%. Thus, both models poorly classified instances of Stage 4, while the ANN model had poor precision and recall for Stage 4. Although ANN had high accuracy with the Normal and Stage 4 classes, low recalls across the rest of the stages lowered its overall performance. The confusion matrices further reiterated that the accurate classification of Stage 4 sarcoidosis was still challenging. Results thus reiterate the need for refining segmentation and feature extraction techniques to gain better classifier performance. This study concludes that while machine learning classifiers show promise for sarcoidosis staging, significant segmentation, and feature extraction improvements are needed to achieve reliable and precise diagnostic outcomes.
|
|
|
|
|
|
|
|
