Dados do Trabalho

Título

A Convolutional Neural Network for the Automatic Quantification of Abnormal Lung Parenchyma Attenuations from Chest Computed Tomography Images

Introdução

Quantitative computed tomography (QCT) techniques might provide objective quantification with some advantages to the visual assessment of abnormal lung parenchyma attenuations.
We aim to evaluate a fully automatically tool (QUAntitative Lung Imaging Tool, QUALIT) for QCT assessment from chest computed tomography (CT) images.

Casuística e Métodos

QUALIT includes two convolution neural network (CNN) that has been trained for automatic lung segmentation and for the classification of low- (emphysema and cysts, LAA), normal- (normal parenchyma, NAA) and high attenuation areas (ground-glass opacities (GGO), crazy paving/linear opacity (CP/LO) and consolidation, HAA). It also includes a densitometry (Dens) tool that computes LAA (-1000 to -950 Hounsfield units, HU), NAA (-949 to - 700 HU), and HAA (-699 to -200 HU). A severity index (SI%) was calculated by adding LAA(%) and HAA(%).
The proposed tool was applied in 806 CT scans (176 normal subjects, 337 emphysema patients, and 293 with interstitial lung disease, ILD). Comparison between QUALIT measurements of abnormal attenuations and pulmonary function tests results were assessed with one-way ANOVA. CNN- and Dens-derived QUALIT measures were also compared and the correlation between them assessed.

Resultados

CNN-SI% and Dens-SI% were very strongly correlated (r = 0.90) and the SI% significantly increased with disease severity (P<0.001). CNN-SI% was higher than Dens-SI% in patients with emphysema and in severe ILD patients (P<0.05). CNN-NAA% was significantly higher than Dens-NAA% in controls (97.9 ± 2 vs 88.6 ± 4.1, P<0.001). CNN-LAA% was significantly lower than Dens-LAA% in Controls (0.4 ± 1.1 vs 2.5 ± 4.5, P<0.001) and in patients with pulmonary emphysema but with normal pulmonary function (1.6 ± 3.4 vs 4.9 ± 4.5, P<0.001) and significantly higher in COPD GOLD III (20.8 ± 18.2 vs 18.5 ± 11.4, P = 0.0453) and GOLD IV subjects (27.9 ± 20.7 vs 21.9 ± 14.2, P<0.001). QUALIT-HAA% was significantly lower than Dens-HAA% in controls (1.6 ± 1.8% vs 8.8 ± 2.1%, P<0.001) and in ILD subjects (29.9 ± 26.7% vs 31.6 ± 19.4%, P<0.001).

Conclusões

QUALIT seems to be a promising tool for the quantification of pulmonary abnormalities allowing the assessment of pulmonary involvement in several lung diseases. Likely, CNN might be used with better performance in less severe cases of both pulmonary emphysema and ILD for the quantification of areas with abnormal lung attenuations close to the density limits defined for normal lung parenchyma.

Palavras Chave

computed tomography, deep learning, densitometry, quantitative computed tomography, artificial intelligence