Machine-Learning-Based Framework for Estimation of Prostate Locations with Anatomical Feature Points on CBCT Images for Image-Guided Target-Based Patient Positioning in Prostate Cancer Radiotherapy

抄録

The aim of our study is to develop a machine-learning (ML) based framework for estimating prostate locations with anatomical feature points (AFPs) on cone-beam computed tomography (CBCT) images for image-guided target-based patient positioning in prostate cancer radiotherapy. Three AFPs were manually determined at a “center” of prostate, a bladder contact point with prostate, and a front point of rectum on each CBCT image. ML architectures, i.e., support vector machine (SVM), artificial neural network (ANN) and random forests (RF), were incorporated into the proposed frameworks. Prostate locations were estimated using seventy-three training sets of distances between each AFP and an average prostate location with reference prostate centroids. The mean locational errors were 1.07 mm in anterior-posterior (AP) and 2.55 mm in superior-inferior (SI) direction using SVM. ANN achieved 2.48 mm in AP and 3.33 mm in SI direction, whereas RF achieved 2.16 mm in AP and 3.03 mm in SI direction, respectively. The proposed framework can be useful for imageguided target-based patient positioning.