Lung tumor segmentation from CT images using 2.5D spatial data and 3D volumetric post-processing

INTRODUCTION: Nowadays, lung cancer is known as one of the serious threats to human health and life globally. To deal with lung cancer, a problem is to accurately segment tumors from CT scans to help doctors in assessment and treatment planning. OBJECTIVES: This study proposes a method for segmenting lung tumors from CT scan data based on exploiting spatial information between slices to feed into the deep learning model and applying 3D volumetric post-processing. In addition, we also provide visualization of 3D tumor reconstruction data. METHODS: We incorporate spatially continuous data into the 2.5D input. In addition, we use a hybrid TransUNet deep learning model, and finally post-process 3D volumes. Our method allows the deep learning model to take advantage of spatial context information along the slices. TransUNet takes advantage of both CNN and Transformer. The 3D volume post-processing technique refines the results, removing discrete noisy blocks. This results in increased continuity and more accurate reconstruction in three-dimensional space. RESULTS: The segmentation results were satisfactory. In case of 2.5D of 3 slices, results on NSCLC-Radiomics reached the Dice over slices of 0.891 ± 0.277, the Dice over 3D volumes of 0.641 ± 0.235, the IoU over slices of 0.875 ± 0.291, the IoU over 3D volumes of 0.508 ± 0.228, False Positives of 94, and False Negatives of 224. In case of 2.5D of 5 slices, results on NSCLC-Radiogenomics reached the Dice over slices of 0.951 ± 0.186, the Dice over 3D volumes of 0.671 ± 0.231, the IoU over slices of 0.941 ± 0.200, the IoU over 3D volumes of 0.539 ± 0.223, False Positives of 8, and False Negatives of 67.. In addition, the 3D tumor visualization showed good visualization capabilities. CONCLUSION: The proposed method has proven to be effective in lung tumor segmentation. In this study, the elements of the proposed method have been analyzed comprehensively and comparatively. In the future, we expect to continue to conduct in-depth research and expand the experiment on patient data in Vietnam.