Fig. 1

The workflow. Models to segment and classify nodules based on their central point coordinates provided by radiologists. CT images were cropped into 64 × 64 × 64 voxel volumes according to the center coordinates and fed into the models. Nodule volumes are passing (1) the 3D UNet for nodule segmentation. With the segmentation mask, surrounding soft tissue (background) can be removed; hence the nodule volume estimation can be performed. Nodule volumes then go through (2) the classification model (3D Attention Net) to predict nodule malignancy and pulmonary fibrosis. Separate datasets (with or without semantic fibrosis information) can be selected as the input to the classifier. The model’s attention at different layers can be visualized and interpreted via attention coefficient maps and CAMs