機械学習を用いた位置照合用CT画像再構成の取り組み

Reconstructing computed tomography for image-guided radiotherapy using machine learning

image-guided radiotherapy（IGRT）は照射前や照射間に撮影された画像を利用して患者の位置や姿勢を修正する技術であり，intensity-modulated radiotherapy（IMRT）や定位照射などの高精度治療を含む大部分の放射線治療において重要である。多くの施設でcone-beam computed tomography（CBCT）やkV-X線画像，MV-X線画像によるX線画像照合が利用されている。

In this study, we evaluated delineated organ contours’ similarity of CT images reconstructed from 2D images and evaluated reconstruction methods. Pseudo CT images were reconstructed from DRR created from the planning CT of 5 breast cancer patients using GAN. The similarity between the planning CT image and the pseudo CT image in the lungs, heart, and thoracic spine was compared using the Dice coefficient. The Dice coefficient for both lungs and the heart were＞0.8, but the thoracic spine was 0.46～0.78. While the organ reproducibility accuracy is low, it was possible to reconstruct the CT image from the 2D image. We will consider reconstructing the virtual CT image from the two-dimensional image for position matching.