Performance Evaluation of Automatic Segmentation based on Deep Learning and Atlas according to CT Image Acquisition Conditions (CT 영상획득 조건에 따른 딥 러닝과 아틀라스 기반의 자동분할 성능 평가)

The study by Jung Hoon Kim evaluates the performance of deep learning (Aview RT ACS, OncoStudio) and atlas-based (Smart Segmentation) automatic segmentation methods under varying CT imaging conditions for lung radiotherapy. Using a Lungman phantom, the research assessed lung volume, Dice similarity coefficient (DSC), and 95% Hausdorff distance (HD) across different tube voltages and currents. Atlas-based methods showed minimal volume variation, while deep learning models, particularly at low tube currents, yielded smaller volumes. Aview RT ACS demonstrated higher DSC and lower HD, indicating superior segmentation accuracy compared to OncoStudio. However, OncoStudio had lower standard deviations, suggesting more consistent results. The study advises caution when using deep learning at low acquisition settings and suggests further research on clinical datasets and image quality optimization. This research contributes to improving automatic segmentation accuracy in radiotherapy by identifying optimal CT acquisition conditions.