Dados do Trabalho

Título

Predicting Pulmonary Function with Screening Chest Computed Tomography

Descrição sucinta do(s) objetivo(s)

This study aimed to develop a deep learning (DL) algorithm to predict pulmonary function metrics, including forced expiratory volume in the first second (FEV1), forced vital capacity (FVC), and the FEV1/FVC ratio using chest computed tomography (CT) images.

Material(is) e método(s)

A Convolutional Neural Network (CNN) was trained on chest CT data from 7,955 participants in the National Lung Screening Trial (NLST). The dataset was split into an 80% development set (6,364 subjects) and a 20% validation test set (1,591 subjects). External validation was conducted on 759 subjects from Brazil, comprising 266 low-dose CT scans for lung cancer screening and 493 conventional CT scans, including patients with emphysema, lymphangioleiomyomatosis, interstitial lung disease, and normal pulmonary function. High-risk classification criteria included FVC and FEV1 predicted values below 80%, and FEV1/FVC ratios outside the range of 70%–90%.
Lung segmentation was performed using the TotalSegmentator tool. The volumetric axial CT data were transformed into coronal projections and divided into three equal sections, resulting in a matrix of size (x, y, 3). These features, along with Total Lung Volume (TLV) and density-based mass and volume measurements for low, normal, and high attenuation areas, were input into the EfficientNetV2m CNN and a Multilayer Perceptron (MLP). The final predictive model integrated both approaches. Model performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), sensitivity (SN), specificity (SP), and balanced accuracy (BA).

Resultados e discussão

In the external validation cohort, the model achieved an accuracy of 88.8% (441/759), with a sensitivity of 91.9% and specificity of 83.5%. The prediction of FEV1, FVC, and the FEV1/FVC ratio demonstrated RMSE variations of ±5.5%, 6.9%, and 11.7%, respectively, with correlation coefficients exceeding 0.7.

Conclusões

The proposed deep learning model demonstrated high accuracy in predicting pulmonary function abnormalities using chest CT scans. This approach has the potential to enhance low-dose CT screening programs by identifying individuals at risk for pulmonary function test (PFT) abnormalities.

Palavras Chave

Artificial intelligence; Pulmonary Function Tests; Chest Computed Tomography

Arquivos