Dados do Trabalho

Título

Estimating Predicted Total Lung Capacity from Chest Computed Tomography Improved Accuracy in Assessing Lung Disease Severity

Descrição sucinta do(s) objetivo(s)

During chest CT scans, patients inhale deeply to approximate total lung capacity (TLC), and lung disease extent is assessed by measuring abnormal parenchymal opacities relative to CT-computed lung volume (CTLV). Disease progression can alter CTLV by increasing or decreasing lung density, necessitating a standardized reference value such as predicted TLC (pTLC), which depends on sex and height. However, height is often missing from medical records, leading to the development of a deep learning-based method to estimate pTLC from chest CT scans (CTpTLC) and improve disease severity assessment.

Material(is) e método(s)

A Convolutional Neural Network (CNN) was trained on chest CT data from 2,455 participants in the National Lung Screening Trial (NLST) who had no radiological abnormalities. The dataset was divided into a 70% development set (1,717 subjects), a 15% validation set (369 subjects), and a 15% external test set (369 subjects). Model performance was assessed using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).
CTLV, CTpTLC, and pTLC were computed for 1,050 subjects with emphysema, 863 with interstitial lung disease (ILD), and 269 with combined emphysema and fibrosis. Lung densitometry classified Low-Attenuation Areas (LAA, ≤ –950 HU), Normal-Attenuation Areas (NAA, –949 to –700 HU), and High-Attenuation Areas (HAA, –699 to –250 HU). A severity index (SI) was calculated as LAA + HAA relative to CTLV, CTpTLC, and pTLC.
The Bland-Altman method evaluated agreement between CTpTLC and pTLC in the test set, and Wilcoxon analysis compared SI relative to CTLV, CTpTLC, and pTLC.

Resultados e discussão

In the test cohort, the CNN model achieved RMSE variations of ±6.1%, an MAE of 4.8%, and a correlation coefficient of 0.85. In subjects with emphysema, ILD and emphysema+ILD, CTLV was significantly lower than both CTpTLC and pTLC (P<0.001), and the difference increased with disease severity significanlty reducing the SI.

Conclusões

The proposed CNN model accurately estimated pTLC, and SI was significantly influenced by the reference volume used for normalization, showing equivalence when expressed relative to CTpTLC or pTLC.

Palavras Chave

Convolutional neural networks; Predicted Total Lung Capacity; Chest Computed Tomography

Arquivos