Bayesian delineation framework of clinical target volumes for prostate cancer radiotherapy using an anatomical-features-based machine learning technique

Our aim was to develop a Bayesian delineation framework of clinical target volumes (CTVs) for prostate cancer radiotherapy using an anatomical-features-based machine learning (AF-ML) technique. Probabilistic atlases (PAs) of the pelvic bone and the CTV were generated from 43 training cases. Translation vectors, which could move the CTV PAs to CTV locations, were estimated using the AF-ML after a bone-based registration between the PAs and planning computed tomography (CT) images. An input vector derived from 11 AF points was fed to three AF-ML techniques (artificial neural network: ANN, random forest: RF, support vector machine: SVM). The AF points were selected from edge points and centroids of anatomical structures around prostate. Reference translation vectors between centroids of CTV PAs and CTVs were given to the AF-ML as teaching data. The CTV regions were extracted by thresholding posterior probabilities produced by using the Bayesian inference with the translated CTV PA and likelihoods of planning CT values. The framework was evaluated based on a leave-one-out test with CTV contours determined by radiation oncologists. Average location errors of CTV PAs along the anterior-posterior and superior-inferior directions without AF-ML were 5.7±4.6 mm and 5.5±4.3 mm, respectively, whereas the errors along the two directions with ANN, which showed the best performance, were 2.4±1.7 mm and 2.2±2.2 mm, respectively. The average Dice's similarity coefficient between reference and estimated CTVs for 44 test cases were 0.81±0.062 with ANN. The framework using AF-ML could accurately estimate CTVs of prostate cancer radiotherapy.