Fully Automated Segment-Anything-Model for Robust Lobe Segmentation in Patients with Interstitial Lung Disease

Sungjoo Lim, Doohyun Park

Abstract: Assessment of lung involvement in computed tomography (CT) images of patients with interstitial lung disease (ILD) is crucial for clinical management. Since the significance of involvement varies across lobes, segmentation of the lobes (left upper, left lower, right upper, right middle, and right lower) is essential. However, deep learning-based lobe segmentation models primarily rely on fissures as key features, which are often not well visualized in CT scans of patients with ILD. Consequently, lobe segmentation models trained on normal lung data often perform poorly on ILD patient data. Moreover, annotating ILD patient data for developing lobe segmentation models is both time-consuming and costly. To address this, we propose a fully supervised approach utilizing the Segment Anything Model (SAM), which has the drawback of requiring manual seed points for target objects. In this study, these seed points are automatically generated based on the output of a lobe segmentation model trained on a normal dataset, using an eroded-contour-based point sampling method. These seed points are then used as inputs for SAM. Additionally, we propose a method to aggregate the output of SAM with single-class segmentation of each lobe for improved robustness. As a result, in three cases with ILD, the dice coefficients of the proposed method showed improvements compared to the original model trained on normal data, from 0.788 to 0.896, 0.786 to 0.869, and 0.795 to 0.845, respectively. Although further validation with larger datasets is needed, this study demonstrates the feasibility of using a fully automated SAM as a pilot study.