Automated segmentation of subsolid pulmonary nodules in CT scans using deep learning

Lung cancer is the second most diagnosed cancer and is the leading cause of cancer-related deaths globally. A pulmonary nodule can turn into cancer, leading to fatal outcomes if left undetected. Compared to other types of pulmonary nodules, subsolid nodules (SSN) pose a higher risk of malignancy. Subsolid nodules can be categorized into two subtypes: ground-glass opacities (GGOs) or part-solid nodules (PSNs). The assessment of SSNs by physicians on cancer risk and the stage is highly dependent on the size and volume of the nodule. Therefore accurate segmentations are crucial for volumetric calculations when dealing with SSNs. Currently, semi-automated methods are deployed to segment the boundaries of SSNs. This requires a radiologist's manual inputs and fine- tuning and could lead to sub-optimal results. Furthermore, there is no study to date which focuses on evaluating the performance of deep learning in SSNs. Over the past decade, deep learning has made significant strides in medical imaging segmentation, and networks like nnUNet have demonstrated great potential in adapting to new datasets. In this research, nnUNet was used to build a fully-automated segmentation model. However, the successful application of the model requires a high-quality dataset with accurate annotations, particularly for the segmentation of SSNs, which has been an area of limited research. To address this, our research focused on creating a carefully curated dataset with annotations provided by an experienced radiologist using a dedicated lung screening workstation. The model achieved a Dice similarity coefficient of 83.3% for the GGOs and 77.6% & 76% for non-solid and solid core respectively for the PSNs on an external validation dataset. The model provides satisfactory segmentation results in a minimal time, without any external input. It is able to learn the behaviour of the semi-automated method to produce similar segmentation. The model has shown promising potential to generate accurate and objective segmentation without human input for the subsolid nodules. The proposed model acts as a good benchmark in the segmentation of subsolid pulmonary nodules.