World Congress Thoracic Imaging June 18-21, 2017, Hynes Convention Center, Boston, Massachusetts June 18-21, 2017, Hynes Convention Center, Boston, Massachusetts

Sponsoring Societies:

Fleischner Society Society of Thoracic Radiology European Society of Thoracic Imaging Japanese Society of Thoracic Radiology Korean Society of Thoracic Radiology
WCTI Home Congress Information Final Program

Back to 2017 Hot Topics

Automatic Classification of Regional Disease Pattern on Volumetric Chest CT of Diffuse Interstitial Lung Disease with Deep Learning Method: Issues on Image Dimensionality
J. Seo, B. Park, S. Lee, N. Kim;
Asan Medical Center, Seoul, KOREA, REPUBLIC OF.

Purpose: To evaluate the effect of image dimension of region of interest (ROI) in developing convolutional neural network (CNN) based automatic classification system of regional patterns on volumetric chest CT with diffuse interstitial lung disease (DILD).
Materials and Methods: From 767 volumetric CT images of DILD, a thoracic radiologist selected cubic 3D region of interest (ROI) and labeled each ROI with following regional patterns; 106 ROIs with normal pattern, 135 ROIs with ground-glass opacity, 160 ROIs with reticular opacity, 126 ROIs with honeycombing, 137 ROIs with emphysema, and 103 ROIs with consolidation. Because each 3D volume has different pixel spacing and size, we resampled various sizes to 30x30x30 images with same size using trilinear interpolation. Resampled data were randomly split into 60% for training, 20% for validation, and 20% for test. To augment the data set, twenty 20x20x20 ROI were generated from each 30x30x30 ROI. Finally, we made axial (2D) images, axial/sagittal/coronal reconstructed (2.5D) images, and three-dimensional (3D) images respectively. We constructed relatively simple CNN architectures having four-convolutional layer, two-pooling layer, and two-dense layer. The basic architectures are totally same but applied with different dimension of convolution filter─totally 3x3 convolution filter for 2D image data, 3x3x3 only first convolution filter for 2.5D image data, and totally 3x3x3 convolution filter for 3D image data. The accuracy in classifying each regional disease patterns were evaluated and compared.
Results: The results show that using 3-dimentional ROD data has higher accuracy of 84.7% than 2.5D (81.7%) and 2D (82.7%) in classifying regional patterns of DILD. We assume accuracy of 2.5D was the lowest due to reduced information, on the contrary, accuracy of 3D having much more information of patterns was the highest.
Conclusions: Using 3-dimentional image data in developing automatic system to classify regional disease pattern of DILD CT based on CNN seems to be more desirable than using axial images, if available. Further study to optimize the network using augmentation methods based on 3D is awaited.

Back to 2017 Hot Topics
Home | About | Contact