Dados do Trabalho

Título

Texture-Based Differentiation Between Emphysema and Airspace Cysts in Chest Computed Tomography

Descrição sucinta do(s) objetivo(s)

To differentiate emphysema from airspace cysts using texture-based convolutional neural networks (CNN) on chest computed tomography (CT) images. Moreover, a fully automated system was implemented to identify abnormal low attenuation areas (LAA) and to assess its quantification and regional distribution.

Material(is) e método(s)

Two CNNs were trained for automatic lung segmentation and classification of low- (emphysema, cysts), normal- (NAAs; normal parenchyma), and high-attenuation areas (HAAs; ground-glass opacities, crazy paving/linear opacity, consolidation). Densitometry also computed LAAs, ≤–950 Hounsfield units (HU), NAAs (–949 to –700 HU), and HAAs (–699 to +50 HU).
CNN estimations of LAA were also subdivided in emphysema and airspace cysts in 343 patients with emphysema and 72 with lymphangioleiomyomatosis (LAM).
Spearman’s correlation coefficient between CNN- and Dens-LAA was calculated, and the predominant pattern (emphysema or airspace cysts) was assessed by CNN in all subjects.

Resultados e discussão

CNN- and Dens-LAA correlated strongly (ρ = 0.79, 0.76 – 0.82, 95% CI).
Both CNN- and Dens-LAA increased with disease severity. CNN were able to differentiate LAA with emphysema from LAA with airspace cysts and, as expected, emphysema was recognized as the predominant LAA pattern in patients with emphysema whereas airspace cysts in patients with LAM.

Conclusões

This study reported both CNN- and Dens-LAA were able to identify LAA and were strongly correlated. However, only CNN texture-based analysis was able to subclassify LAA related to emphysema from LAA related to airspace cysts.

Palavras Chave

Airspace cysts; Emphysema; Texture-based convolutional neural network

Arquivos

Área

Tórax