Chest X-Ray Enhancement Based On Bone-Suppression Model Using Multi-Dilated-Rate Strategy
Z Chen*, H Sun, L Song, G Ren, J Yang, J Cai, The Hong Kong Polytechnic University, Hong Kong
Presentations
PO-GePV-I-85 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To develop a bone-suppression model based on multi-dilated-rate strategy that improves the texture feature extraction of lung region in chest X-ray (CXR) images.The enhanced CXR images aim to assist radiologists in making a more accurate diagnosis of pneumonia diseases.
Methods: The JSRT dataset of 217 CXR cases were adopted for the model development. A number of 177 cases were randomly selected for model training and validation using eight-fold cross-validation. The main framework of the model was a network composed of ResNet. Each ConvBlock consisted of convolutional layers with multi dilated rates. The dilated operation was used to reduce the amount of calculation and ensure that each element in the deep feature map can be fully captured and make full use of the shallow features. The skip connections fused the shallow information of the image with the deep information in the form of element-wise, thereby retaining more lung texture information, and suppressing the bone signal in the CXR images. In the testing stage, the remaining 40 cases were used to verify the performance of the proposed model in terms of anatomy. The quantitative evaluation metrics included peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean squared error (RMSE).
Results: Compared with the current bone suppression models, the enhanced CXR images had relatively better visual effects of bone suppression, and the obtained images were comparable to the target CXR images in numerical results. The PSNR between them was 38.22±3.66 dB, the SSIM was 0.9930±0.0033, and the RMSE was 0.0136±0.0079. In the results of paired t-test, the differences were all statistically significant.
Conclusion: The multi-dilated-rate strategy based CXR enhancement model is able to enhance the lung feature. It holds great promise to help clinical radiologists to diagnose pneumonia diagnosis of disease.
Funding Support, Disclosures, and Conflict of Interest: HMRF COVID190211
Keywords
Taxonomy
Contact Email