Data

To develop the artificial intelligence (AI) algorithm for this study, we utilized a comprehensive CT chest dataset consisting of 1,500 cases. This dataset includes 23 labels, normal lung appearance, honeycombing, fibrosis, pulmonary edema, septal thickening, pleural effusion, reticulation, air-trapping, pneumothorax, emphysema, traction bronchiectasis, consolidation, atelectasis, lesions, mucous plugging, ground-glass opacity, peri-bronchial vascular thickening, pleural thickening, cavitation, centrilobular nodule (tree-in-bud appearance), calcification, cysts, and infiltrates.  Among these, reticulation, honeycombing, traction bronchiectasis, ground-glass opacities, cysts, septal thickening, and fibrosis are classified as interstitial lung abnormalities (ILAs), as highlighted in the Fleischner Society’s position paper [1]. Detecting and categorizing these abnormalities is crucial, as they can indicate early signs of progressive lung diseases, including interstitial lung diseases (ILDs).

A few example cases are shown in Figure 1 highlighting different chest abnormalities.

• A chest CT image with consolidation in the right upper lobe and ground-glass opacity in the left upper lobe is shown in Figure 1(a).
• Figure 1(b) presents a case of emphysema in both lower lobes.
• A case of honeycombing in the left upper lobe and ground-glass opacity in the right upper lobe is presented in Figure 1(c).
• Figure 1(d) shows a chest CT image with ground-glass opacity in both lungs

Model training

For this study, we employed the 3D ResNet18 deep learning model, which was trained and validated using this dataset. The dataset was divided into training (1200 cases) and test sets (300 cases). The model was trained on the training set and evaluated on the test set. 

To ensure consistency in analysis, the dataset is pre-processed with standardized spacing, scale, and resolution across all images. The model predicts the probability of each abnormality, allowing for an accurate estimation of whether a particular abnormality is present or absent in a given chest CT scan image.  

We have incorporated this study into our upcoming product, mCT, which focuses on lung nodule detection and related findings. As part of this integration, the system is designed to generate comprehensive reports that include both lung nodule and ILA findings. Currently, our product supports eight key labels: normal, reticulation, atelectasis, bronchiectasis, emphysema, pleural effusion, calcification, and consolidation.  

Model evaluation

To evaluate the performance of our proposed approach, we employed accuracy and area under the receiver operating characteristic (AUROC) as evaluation metrics. Accuracy measures the proportion of correctly classified cases, providing the model’s predictive performance. Meanwhile, AUROC evaluates the model’s ability to distinguish between different conditions by analyzing the trade-off between sensitivity and specificity across various threshold settings. A higher AUROC score indicates better discriminatory power, ensuring that abnormalities are correctly identified while minimizing false positives and false negatives.