Structure optimization of deep convolutional neural network for automatic classification of calcifications and stents in coronary computed tomography angiography

IR Web Site Open Access

Bibliographic Information

Other Title
  • 冠動脈 CT における石灰化とステントの自動分類のためのDCNN 構造最適化
  • カンドウミャク CT ニ オケル セッカイカ ト ステント ノ ジドウ ブンルイ ノ タメ ノ DCNN コウゾウ サイテキカ

Search this article

Description

Purpose ： The purpose of this study was to classify coronary arteries with calcifications or stents and normal coronary arteries with high accuracy as a computer-aided diagnosis system in coronary computed tomography angiography (CCTA). Methods ： Altogether, 49 CCTA scans were taken retrospectively. We obtained 72,051 cross-sectional images of coronary arteries (13,035 calcified images, 14,382 stent images, and 44,634 normal images) among CCTA images. For the deep convolutional neural network (DCNN), we used VGG-22 with six additional convolutional layers from the publicly available VGG-16. From VGG-22, the number of intermediate layers in the full connected layers were increased from three to five. On the target images, 64,846 images were used as training data, and 7,205 images were used as test images. We performed hold-out validation at seven times, calculated accuracy and positive predictive value of each group, and compared the three DCNNs. Results ： The positive predictive value of the calcified group was 95.3% for 3 full connected layers, 95.6% for 4 layers, and 95.9% for 5 layers. The positive predictive value of the stent group was 99.2% for 3 layers, 98.9% for 4 layers, and 99.3% for 5 layers. Conclusion ： The accuracy of automatic classification with calcifications and stents using CCTA can be improved by increasing the fully connected layer of DCNN. The results show that DCNN is effective for triage of coronary artery disease using CCTA.

Journal

Related Projects

See more

Keywords

Details 詳細情報について

Report a problem

Back to top