능동적 학습과 설계 프로그래밍을 이용하여 효율적인 3D프린팅을 위한 반자동 작업 과정 및 임상적 유용성

Abstract

In the field of medicine, 3D printing (3DP) technology has been applied to patient-specific surgical guides, simulators, surgical planning, education, implants. The essential tasks such as image acquisition, segmentation, 3D computer aided design (CAD) modeling, or measurements must be performed in medical field. However, these tasks can be repetitive, time-consuming, labor-intensive, and lack of consistency. To address these shortcomings, improvements can be achieved by segmenting CT images using deep learning, generating a 3D model, and performing semi-automated modeling using a script-based application programming interface (API). This study was conducted in three parts: 1) Active learning (AL) was utilized to automated and enhanced segmentation, reducing the labeling workload, 2) automated patient-specific modeling and measurement of landmarks were achieved through the 3D models based on automated segmentation using script-based application programming interfaces (API), and 3) the usefulness and efficiency of patient-specific surgical guides produced using 3DP technology were demonstrated with clinical application. Firstly, medical image segmentation is essential to obtain various information within the human body, providing visualization of anatomical structures and information on diagnosis, surgical planning, organs, or lesions for medical professionals. Conventional image segmentation often involves manual segmentation on a pixel-by-pixel basis using various tools. However, these segmentation methods often face difficulties in achieving consistent segmentation due to various factors such as low contrast and image noise, leading to significant time consumption. In recent years, significant advances in image segmentation have been made using deep learning models such as convolution neural networks (CNN), fully convolutional networks (FCN), and U-Net. In this study, we introduced an AL approach for enhanced segmentation using a smart labeling, which efficiently increases labeled data by training on a small initial dataset with manually labeled data and correcting the predicted new data with human experts in kidney CT with renal cell carcinoma, mandibular condyle CBCT, thracoabdominal aortic dissection CT, abdominal aortic aneurysm CT. Various networks including 2D or 3D U-Net, Cascade 3D U-Net, UNETR, SwinUNETR, and nnU-Net were selectively used for AL in various tasks. The evaluation was performed using dice similarity coefficient (DSC) and Hausdorff distance (HD) to assess the agreement of areas and distances at predefined stages, accuracy for the multi-classes to be segmented, time comparison for segmentation using manual segmentation and smart labeling, and stress tests to determine the optimal amount of data. Secondly, using a 3D CAD tool based on a 3D model obtained through automated segmentation, patient-specific modeling or landmark measurement for diagnosis and lesion tracking can be performed. By using CAD systems, design modifications can be easily generated to improve accuracy in clinical applications, and they are optimized to meet specific requirements of clinicians. However, like manual segmentation, conventional modeling methods are time-consuming and require significant labor. An automated CAD modeling system using a script-based API provides excellent performance, accuracy, and efficiency in a short time, overcoming the limitations of conventional modeling methods. Automated CAD modeling typically involves setting inputs and parameters for 3D modeling, generating algorithm and programming code based on the specifications, and integrating the code with API of CAD software, enabling modeling and modification. In thoracoabdominal aortic dissection CT and abdominal aortic aneurysm CT, automated 3D CAD modeling and measurement were compared with conventional manual methods, and a series of processes required for automated modeling were optimized, modularized, and validated to perform design and measurement for patients with various anatomical structures. The evaluation involved analyzing the accuracy of 3D models and measurements obtained through conventional manual and automated CAD modeling, as well as the time required for modeling. Finally, 3DP technology was found to have the advantage of addressing clinical unmet needs for pediatric cases, rare and complex conditions, and surgeries that are difficult to standardize, as most medical devices are commonly developed or customized for adult patients with common diseases. 3DP technologies proceed by adding materials layer-by-layer until the object is completely built. It can be classified into various methods depending on the materials used and the printing process. The 3D model obtained through automated segmentation and 3D CAD modeling is converted to standard tessellation language (STL) format and printed using the appropriate 3DP, followed by sterilization before clinical application. We introduce two new reconstruction techniques using patient-specific 3D-printed graft reconstruction guides in open surgical repair of thoracoabdominal aortic dissection: (1) model-based technique (MBT) that presents the projected aortic graft, visualizing the main aortic body and its major branches and (2) guide-based technique (GBT) in which the branching vessels in the visualizing guide are replaced by marking points identifiable by tactile sense. The effectiveness was demonstrated by evaluating conventional and new techniques base on accuracy, marking time requirement, reproducibility, and results of survey to surgeons on the efficiency and efficacy. The accuracy of 3DP guides can be affected by various factors such as the external environment, duration of 3D printer, resolution, materials, shape of the 3D model, printing conditions, and post-processing, which may lead to discrepancies between the printed model and the original 3D model. In conclusion, we have developed an automated workflow for the segmentation and modeling processes necessary for applying 3DP in the clinical application. This approach significantly reduces repetitive and labor-intensive tasks, maintains consistency through codification, and saves time, thereby alleviating their workflow and accelerating the application of 3DP compared to conventional methods. Furthermore, the clinical usefulness and efficiency of 3D printed patient-specific surgical guides demonstrated in addressing clinical unmet needs.