Superior performance of three-dimensional to two-dimensional convolutional neural network for predicting airflow...

Chronic obstructive pulmonary disease (COPD) may be inconsistent with the severity of airflow limitation. This causes COPD underdiagnosis, necessitating approaches that facilitate timely diagnosis and intervention. Combining deep learning models (based on medical imaging) with regression methods improves numerical functional predictions. We aimed to evaluate and compare the prediction performance of two deep learning-based models (two-dimensional [2D]-convolutional neural network (CNN) and three-dimensional [3D]-CNN) for the percentage predicted forced expiratory volume in 1 s (%FEV₁) in patients with COPD.

ResNet18-based regression prediction models were constructed for %FEV₁ based on 200 computed tomography (CT) datasets. Five-fold cross-validation was performed to develop the predictive models, which were externally validated using 20 data points. In addition, 200 internal CT datasets were assessed using commercial software to develop a regression model for predicting airway (% wall area) and parenchymal indices (% low-attenuation volume).

The 3D-CNN model demonstrated superior performance with an average root mean squared error (RMSE) of 10.73 and a correlation coefficient of 0.88, compared with that of the 2D-CNN model (RMSE: 16.76, correlation coefficient: 0.66) during internal validation. In the external validation approach, the 3D-CNN model maintained a performance (RMSE: 11.48, correlation coefficient: 0.59) better than that of the 2D-CNN model (RMSE: 12.38, correlation coefficient: 0.47), with both models outperforming the commercial software analysis (RMSE: 23.18).

Volumetric analysis using 3D-CNN may sufficiently capture the complex structural features of COPD in CT images. Further studies are required to validate these models with larger datasets and determine their validity for longitudinal applications.