Evaluating accuracy in artificial intelligence-powered serial segmentation for sectional images applied to morphological studies with three-dimensional reconstruction.

Abstract

Three-dimensional (3D) reconstruction is time consuming owing to segmentation work. We evaluated the accuracy of the artificial intelligence (AI)-based segmentation and tracking model SAM-Track for segmentation of anatomical or histological structures and explored the potential of AI to enhance research efficiency. Images (obtained via computed tomography [CT] and magnetic resonance imaging [MRI]), anatomical sections from a Visible Korean Human open resource, and serial histological section images of cadavers were obtained. Six structures in the CT, MRI, and anatomical sections and seven in the histological sections were segmented using SAM-Track and compared with manual segmentation by calculating the Dice similarity coefficient. Segmented images were then reconstructed three dimensionally. The average Dice scores of CT and MRI results varied (0.13-0.83); anatomical sections showed mostly good accuracy (0.31-0.82). Clear-edged structures, such as the femur and liver, had high scores (0.69-0.83). In contrast, soft tissue structures, such as the rectus femoris and stomach, had variable accuracy (0.38-0.82). Histological sections showed high accuracy, especially for well-delineated tissues, such as the tibia and pancreas (0.95, 0.90). However, the tracking of branching structures, such as arteries and veins, was less successful (0.72, 0.52). In 3D reconstruction, high Dice scores were associated with accurate shapes, whereas low scores indicated discrepancies between the predicted and true shapes. AI-based automatic segmentation using SAM-Track provides moderate-to-good accuracy for anatomical and histological structures and is beneficial for conducting morphological studies involving 3D reconstruction.