Unsupervised 3D Lung Segmentation by Leveraging 2D Segment Anything Model

Abstract

Lung segmentation is the first important step for lung nodule detection and lung cancer analysis. Deep neural networks have achieved state-of-the-art for most tasks in medical image analysis, including lung segmentation. However, training a deep learning model requires a large amount of annotated samples, which is not practical in medical imaging. In this study, we make efforts to perform unsupervised lung segmentation on 3D lung CT data by leveraging foundational 2D segment anything model (SAM). The approach utilizes SAM to segment 2D slides and generate 2D masks, then reconstruct multiple 2D masks from the same subject into one 3D mask. In such a way, we can train a 3D lung segmentation model by using the reconstructed 3D masks without the requirement of any ground truth annotations, namely, in an unsupervised manner. The evaluation on LUNA16 dataset shows our proposed unsupervised 3D model achieves comparable results with enhanced stability compared to the supervised one trained with ground truth annotations.