Abstract Rationale Diffusing capacity of the lung for carbon monoxide (DLCO) is a key marker of gas-exchange impairment, yet it requires specialized equipment that may be unavailable in certain areas. We investigated whether 2D lung computed tomography (CT) image projections can predict measured DLCO, enabling image-based estimation when pulmonary testing is missing or unavailable. Methods CT SHARP Inspiratory images from n = 2,958 COPDGene participants with DLCO measurments were used to generate 2D coronal lung projections. The lungs were segmented to suppress extrathoracic signal, and images were normalized per subject. Projections were resized to a fixed dimension for downstream prediction modeling. The dataset was randomly split into 70:15:15 (training:testing:validation). We trained convolutional neural networks (CNNs) to regress raw DLCO using Mean Squared Error (MSE) as the loss function. We used EfficientNetB0 and EfficientNetV2s at different pre-training and augmentation strategies. Evaluation metrics were MAE, RMSE, R², Pearson r, and Spearman r. Results The best performing image-based prediction model was the pretrained EfficientNetV2s with radiology-oriented augmentation (xrayaug) (Table 1; MAE: 3.48, RMSE: 4.43, R²: 0.61, Pearson r: 0.78, Spearman r: 0.77; p 0.001), which outperformed non-pretrained and non-augmented baselines. Pretraining provided the largest gains, while xrayaug augmentation yielded additional improvements. Interestingly, the trivialaug wide and the randaug do not seem to improve over the non-augmented EfficientNetV2s model (R2: 0.56, 0.55 compared to 0.58). Furthermore, EfficientNetB0 seems to perform better at lower epochs than EfficientNetV2s, an advantage which is lost at higher epochs. Conclusions Our study showed that 2D lung CT projections can estimate raw DLCO with good accuracy. These findings suggest that projection-level image features encode gas-exchange-relevant information and could help towards DLCO imputation when DLCO tests are unavailable, or when retrospective datasets have missing physiology, or longitudinal monitoring is incomplete. Funding This work was supported by NHLBI R01HL159805. The COPDGene study (NCT00608764) is supported by grants from the NHLBI (U01HL089897 and U01HL089856), by NIH contract 75N92023D00011, and by the COPD Foundation through contributions made to an Industry Advisory Committee that has included AstraZeneca, Bayer Pharmaceuticals, Boehringer-Ingelheim, Genentech, GlaxoSmithKline, Novartis, Pfizer and Sunovion. This abstract is funded by: NHLBI R01HL159805, U01HL089897, U01HL089856, NIH contract 75N92023D00011
Kee et al. (Fri,) studied this question.