IMPROVING CHEST PATHOLOGIES DETECTION FROM CHEST X-RAY WITH DEEP LEARNING USING TRANSFER LEARNING AND IMAGE ENHANCEMENT
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Srinakharinwirot University
Abstract
This research is concerned with chest radiography, which is essential for doctors to determine and follow up on lung disease. However, practicing radiologists have an insufficient ability to identify diseases in chest x-ray images. Therefore, the researchers developed deep-learning models to mitigate this problem, and CheXNet is one of the state-of-the-art models that can detect 14 lung pathologies. This research applied six image enhancement techniques to the x-ray images before using ChexNet to improve detection performance. The six techniques consisted of Gamma, Complement, HE, CLAHE, BCET, and MMCS. In addition, we studied the effectiveness of using a single enhancement technique (single channel) and a combination of them to the original image (multi-channel). Gamma gave the highest and most stable detection improvement using a single enhancement technique at 0.628% AUCROC in 14 diseases. Combining the original image, Gamma-enhanced image, and CLAHE-enhanced image shows 0.7% AUCROC improvement for 14 diseases. Moreover, this combination offers outstanding Pneumonia detection, which is 2% more than CheXNet.
This research is concerned with chest radiography, which is essential for doctors to determine and follow up on lung disease. However, practicing radiologists have an insufficient ability to identify diseases in chest x-ray images. Therefore, the researchers developed deep-learning models to mitigate this problem, and CheXNet is one of the state-of-the-art models that can detect 14 lung pathologies. This research applied six image enhancement techniques to the x-ray images before using ChexNet to improve detection performance. The six techniques consisted of Gamma, Complement, HE, CLAHE, BCET, and MMCS. In addition, we studied the effectiveness of using a single enhancement technique (single channel) and a combination of them to the original image (multi-channel). Gamma gave the highest and most stable detection improvement using a single enhancement technique at 0.628% AUCROC in 14 diseases. Combining the original image, Gamma-enhanced image, and CLAHE-enhanced image shows 0.7% AUCROC improvement for 14 diseases. Moreover, this combination offers outstanding Pneumonia detection, which is 2% more than CheXNet.
This research is concerned with chest radiography, which is essential for doctors to determine and follow up on lung disease. However, practicing radiologists have an insufficient ability to identify diseases in chest x-ray images. Therefore, the researchers developed deep-learning models to mitigate this problem, and CheXNet is one of the state-of-the-art models that can detect 14 lung pathologies. This research applied six image enhancement techniques to the x-ray images before using ChexNet to improve detection performance. The six techniques consisted of Gamma, Complement, HE, CLAHE, BCET, and MMCS. In addition, we studied the effectiveness of using a single enhancement technique (single channel) and a combination of them to the original image (multi-channel). Gamma gave the highest and most stable detection improvement using a single enhancement technique at 0.628% AUCROC in 14 diseases. Combining the original image, Gamma-enhanced image, and CLAHE-enhanced image shows 0.7% AUCROC improvement for 14 diseases. Moreover, this combination offers outstanding Pneumonia detection, which is 2% more than CheXNet.