Enhancing Pulmonary Involvement Assessment in Computed Tomography Scans Using Predictive Models Based on Demographics and Anatomical Measurements
Principal Investigator
Name
Alysson Carvalho
Degrees
M.D., Ph.D.
Institution
ID'Or
Position Title
Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1222
Initial CDAS Request Approval
Apr 2, 2024
Title
Enhancing Pulmonary Involvement Assessment in Computed Tomography Scans Using Predictive Models Based on Demographics and Anatomical Measurements
Summary
Objective:
This study focuses on enhancing the assessment of pulmonary involvement in computed tomography (CT) scans, commonly referred as the volume of abnormal lung opacities adjusted to CT-computed lung volume (CTLV). However, lung diseases modify lung parenchyma structure, affecting total lung capacity and CTLV, thereby necessitating improved accuracy in measuring lung involvement. We propose a methodology to calculate predicted CTLV (pCTLV) using demographic factors (sex, age, body weight, height) and anatomical measurements (maximum lengths of clavicle, scapula, sternum) derived from chest CT scans. This is particularly useful as patient height is often not recorded in routine CT studies.
Methods:
We employed a U-Net Convolutional Neural Network (CNN) for automatic lung segmentation from 173 CT scans of healthy individuals. CTLV was computed using pixel dimensions and slice thickness. The bilateral clavicles, scapulae, and sternum were automatically segmented, and their maximal lengths were measured. A Support Vector Regression (SVR) model was developed using body height, weight, sex, and age to estimate pCTLV. An alternative pCTLV model (pCTLV’) was also formulated, substituting body height with the maximal lengths of clavicles, scapulae, and sternum. The dataset was randomly split into 70% training and 30% testing, and we assessed various hyperparameter ranges for the SVR model with the "Radial Basis Function" kernel. The model's accuracy was already evaluated using a Bland-Altman plot.
Initial Results:
The pCTLV model demonstrated a mean absolute error of 334.2 ml in training, 470.0 ml in testing, and a global error of 385.5 ml (R2 Training = 0.71, Test = 0.63, Global = 0.68). The pCTLV’ model showed a mean absolute error of 341.2 ml in training, 508.7 ml in testing, and a global error of 380.1 ml (R2 Training = 0.74, Test = 0.60, Global = 0.71). We observed a 14.1 ml and -8.4 ml specific bias between CTLV, pCTLV, and pCTLV’, respectively and this bias increased with CTLV.
Future Perspectives with the NLST Database:
To assess the specific bias between CTLV, pCTLV, and pCTLV’ in the calculation of the extent of pulmonary involvement, we aim to assess the National Lung Screening Trial (NLST) dataset. From each CT, the same U-Net Convolutional Neural Network (CNN) will segment the lung parenchyma and another, CNN trained for the automatic classification of low- (LAAs; emphysema, cysts), normal- (NAAs; normal parenchyma), and high-attenuation areas (HAAs; ground-glass opacities, crazy paving/linear opacity, consolidation) will be applied. The amount of abnormal lung parenchyma attenuation (AA) will be computed as AA (ml) = LAA + HAA and will be referred to CTLV (AA_CTLV), pCTLV (AA_pCTLV) and pCTLV’ (AA_ pCTLV’). The specific bias between AA_CTLV and AA_pCTLV, AA_CTLV and AA_ pCTLV’, and between AA_pCTLV and AA_ pCTLV’ will also be evaluated using a Bland-Altman plot.
Expected results:
We expect that the bias between AA_CTLV and AA_pCTLV and between AA_CTLV and AA_ pCTLV’ increased with disease severity.
This study focuses on enhancing the assessment of pulmonary involvement in computed tomography (CT) scans, commonly referred as the volume of abnormal lung opacities adjusted to CT-computed lung volume (CTLV). However, lung diseases modify lung parenchyma structure, affecting total lung capacity and CTLV, thereby necessitating improved accuracy in measuring lung involvement. We propose a methodology to calculate predicted CTLV (pCTLV) using demographic factors (sex, age, body weight, height) and anatomical measurements (maximum lengths of clavicle, scapula, sternum) derived from chest CT scans. This is particularly useful as patient height is often not recorded in routine CT studies.
Methods:
We employed a U-Net Convolutional Neural Network (CNN) for automatic lung segmentation from 173 CT scans of healthy individuals. CTLV was computed using pixel dimensions and slice thickness. The bilateral clavicles, scapulae, and sternum were automatically segmented, and their maximal lengths were measured. A Support Vector Regression (SVR) model was developed using body height, weight, sex, and age to estimate pCTLV. An alternative pCTLV model (pCTLV’) was also formulated, substituting body height with the maximal lengths of clavicles, scapulae, and sternum. The dataset was randomly split into 70% training and 30% testing, and we assessed various hyperparameter ranges for the SVR model with the "Radial Basis Function" kernel. The model's accuracy was already evaluated using a Bland-Altman plot.
Initial Results:
The pCTLV model demonstrated a mean absolute error of 334.2 ml in training, 470.0 ml in testing, and a global error of 385.5 ml (R2 Training = 0.71, Test = 0.63, Global = 0.68). The pCTLV’ model showed a mean absolute error of 341.2 ml in training, 508.7 ml in testing, and a global error of 380.1 ml (R2 Training = 0.74, Test = 0.60, Global = 0.71). We observed a 14.1 ml and -8.4 ml specific bias between CTLV, pCTLV, and pCTLV’, respectively and this bias increased with CTLV.
Future Perspectives with the NLST Database:
To assess the specific bias between CTLV, pCTLV, and pCTLV’ in the calculation of the extent of pulmonary involvement, we aim to assess the National Lung Screening Trial (NLST) dataset. From each CT, the same U-Net Convolutional Neural Network (CNN) will segment the lung parenchyma and another, CNN trained for the automatic classification of low- (LAAs; emphysema, cysts), normal- (NAAs; normal parenchyma), and high-attenuation areas (HAAs; ground-glass opacities, crazy paving/linear opacity, consolidation) will be applied. The amount of abnormal lung parenchyma attenuation (AA) will be computed as AA (ml) = LAA + HAA and will be referred to CTLV (AA_CTLV), pCTLV (AA_pCTLV) and pCTLV’ (AA_ pCTLV’). The specific bias between AA_CTLV and AA_pCTLV, AA_CTLV and AA_ pCTLV’, and between AA_pCTLV and AA_ pCTLV’ will also be evaluated using a Bland-Altman plot.
Expected results:
We expect that the bias between AA_CTLV and AA_pCTLV and between AA_CTLV and AA_ pCTLV’ increased with disease severity.
Aims
- We propose a methodology to calculate predicted CTLV (pCTLV) using demographic factors (sex, age, body weight, height) and anatomical measurements (maximum lengths of clavicle, scapula, sternum) derived from chest CT scans. This will be particularly useful as patient height is often not recorded in routine CT studies.
- To assess difference between the amount of abnormal lung parenchyma attenuation (AA) referred to CTLV, pCTLV and pCTLV’.
Collaborators
Alan Ranieri
Bruno Hochhegger
Rosona Rodrigues
Errison Alves