Dados do Trabalho

Título

Predicting Pulmonary Function Abnormalities with a Texture-Based Quantification of Normal Lung Parenchyma in Chest Computed Tomography

Descrição sucinta do(s) objetivo(s)

To predict pulmonary function abnormalities (restrictive and/or obstructive patterns) with the texture-based convolutional neural networks (CNN) quantification of normal lung index (NLI) on conventional and low-dosage chest CT images.

Material(is) e método(s)

Two CNNs were trained for automatic lung segmentation and classification of low- (LAAs; emphysema, cysts), normal- (NAAs; normal parenchyma), and high-attenuation areas (HAAs; ground-glass opacities, crazy paving/linear opacity, consolidation). NLI was calculated as NLI=100x[NAA/(LAA + NAA + HAA)]. We assessed 880 conventional CT scans from 176 normal subjects, 343 patients with emphysema, 68 with lymphangioleiomyomatosis (LAM) and 293 patients with interstitial lung disease (ILD). A Receiver Operating Characteristic (ROC) analysis was used to assess the performance of NLI to distinguish controls from emphysema, LAM and ILD patients with abnormal pulmonary function tests (PFTs). The criteria used to define normal spirometry findings were a prebronchodilator percent of FVC exhaled in first second (hereafter, FEV1/FVC) greater than or equal to 70% and FVC% and FEV1% values greater than or equal to 80%.

Resultados e discussão

Out of 880 subjects, 161 controls, 86 patients with emphysema, 34 LAM and 77 ILD subjects were considered as having normal PFTs. NLI 5, 50 and 95% percentiles in healthy control subjects were 93.1, 99.6 and 99.9%, respectively. A reference equation for NLI was also generated: NLI (mL) = 842*Sex (1 = male, 0 = female) + 4164 * Height (m) - 2451 (R2 = 0.55, adjusted R2 = 0.55, F-statistic = 101 and P < 0.0001). NLI expressed as a percentage of predicted values decreased with disease severity. The NLI threshold of 96.1% (AUC = 0.96, 95%CI 0.94–0.99, sensitivity of 0.91 and specificity of 0.88, accuracy of 0,88 and F1-Score of 0,75) was able to differentiate controls from patients with abnormal PFTs.

Conclusões

This study reported reference values, thresholds, and reference equations for NLI derived from quantitative CT scans assessments of subjects with normal lung function and CT findings. NLI estimations might aid in the screening of patients with lung parenchymal abnormalities.

Palavras Chave

Quantitative computed tomography; Texture-based convolutional neural network; Artificial intelligence

Arquivos

Área

Tórax