Program Information
Improving Lung Visibility in Chest Radiography by Suppressing Rib Bone with Deep Learning
D Lee1, H Kim1* , (1)Department of Radiation Convergence Engineering, College of Health Science, Yonsei University, 1 Yonseidae-gil, Wonju, Gangwon, 220-710, Korea
Presentations
SU-I-GPD-J-29 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Although dual energy chest radiography improved visibility of lung field by decomposing bony structures in thorax cage, it delivered relatively high radiation dose compared to conventional chest radiography due to additional exposure with different energy spectra. Therefore in this study, we proposed alternative method for providing bone suppression chest radiography by using image processing technique with deep learning.
Methods: To train from conventional chest radiography to dual energy bone decomposed images with deep learning, we modeled convolution neural network with 5 hidden layers. Input data for training model are sub regions extracted from chest radiography and pixel values of dual energy tissue images are used as output values for the corresponding sub regions. The number and size of total training sub regions are 38,924 and 8 × 8. The number of hidden layer was 64,128,256,128 and 64. The learning rate was 0.0001. After training process, we evaluated model with untrained chest radiography which acquired with different beam quality and exposure conditions.
Results: The results of our training model converged with absolute error of 0.154 and showed 100 % accuracy with training data. Although there are several image noises in results, deep learning showed possibility of decomposing bony structure in chest radiography without additional exposure.
Conclusion: In conclusion deep learning, which is hot issue in medical image, improved visibility of lung field by decomposing bony structure without any additional radiation exposure. However our study used only one chest radiography as training data thus it is necessary to train more chest radiography in order to improve image quality and generalize our suggested image processing technique.
Contact Email: