Purpose: The precise prostate auto-setup in the cone-beam CT (CBCT) guided radiotherapy is complicated by two factors: (1) the low image contrast of the marker-free prostate; (2) and the variations between the planning CT (pCT) and CBCT acquired at the different time. To address these problems, we propose a novel patient-specific residual network strategy to achieve Six-Degrees-of-freedom (6D) auto-setup for the marker-free prostate.
Methods: The proposed method includes four parts: (1) Training data generation: an image augmentation model was applied to the pCT to generate the training dataset with the corresponding prostate bounding and landmarks. (2) Network training: we developed a two-step task-based residual network (T2RN) to localize the prostate and then detect the landmarks inside the localized region. (3) Prediction on the CBCT: an iterative filter scheme was presented to find a transformation to the CBCT so that the gray value distribution can match well with the training image. (4) Determination of the 6D shift: The translation and rotation errors in CBCT were determined using the transformation matrix from the landmarks in the pCT.
Results: For the clinical study, the evaluations were carried on the 80 cases of CBCT, which were augmented from the patient with 20 cases of CBCT acquired at a different time. The mean and standard deviation errors in the anterior-posterior, left-right, superior-inferior, yaw, pitch, and roll were 0.64Â±1.40 mm, 0.15Â±1.28 mm, -0.46Â±1.17 mm, 0.21ÂºÂ±0.60Âº, -0.61ÂºÂ±0.48Âº, and 0.23ÂºÂ±0.44Âº, respectively. The correlation coefficients are 0.99, 0.99, 0.99, 0.91, 0.94, and 0.95. The prediction time for the per case of CBCT was less than 0.5s.
Conclusion: Prostate setup has been mainly investigated so far using the image registration which suffers from either low accuracy or long processing time. T2RN presents a very promising alternative which can offer simultaneously high precision and fast processing for marker-free prostate 6D auto-setup.