CONVOLUTIONAL NEURAL NETWORKS FOR SEMANTIC SEGMENTATION IN 3D RECONSTRUCTION PROCESS OF THE PULMONARY ACINUS

The pulmonary acinus, encompassing respiratory bronchioles, alveolar ducts, alveoli, alveolar sacs, and their associated blood vessels, serves as the main unit for gas exchange in the lung. To accurately characterize fluid micromechanics within this complex 3D acinar structure and the corresponding airflow, 3D reconstruction of the lung parenchyma using computational fluid dynamics is crucial. Essential to this process is the precise segmentation of stacked 2D images, forming the basis for a reliable 3D model. Convolutional neural networks (CNNs) have shown remarkable success in tasks such as image classification, semantic segmentation, and object detection, and are particularly effective in medical image segmentation. This study explores the efficacy of several advanced deep convolutional architectures and their variants for segmenting the pulmonary acinus lung field. The models were assessed using rat lung images acquired through synchrotron radiation-based X-ray tomographic microscopy (SRXTM). Balancing accuracy and computational efficiency, our results demonstrate that SegNet (trained from scratch) and pretrained PSPNet are the most suitable for this type of imagery, achieving mean Intersection over Union (mIoU) values of 0.85 and 0.9, respectively. Ultimately, these segmented images facilitated the complete 3D reconstruction of the pulmonary acinus.