Systematic Study of Patient-Specific Organs at Risk Auto-Segmentation on Daily kVCT Images for Adaptive Head and Neck Radiotherapy

Abstract

<h3>Purpose/Objective(s)</h3> A newly developed biology-guided radiotherapy (BgRT) system has been set up in our institution, which enables high-quality kVCT imaging for each treatment fraction. This study explored deep learning-based patient-specific auto-segmentation to facilitate adaptive radiotherapy, based on data of the first head and neck patient case treated with this new system. <h3>Materials/Methods</h3> The esophagus, larynx, and pharynx were selected as the organs at risk (OAR) for auto-segmentation analysis. A population network was first learned on a population dataset, which contains 67 different patient cases from conventional intensity-modulated radiotherapy (IMRT). Then the pre-trained population network was adapted to this specific patient using a transfer learning method based on the longitudinal data from the initial treatment planning and 17 sets of daily image guidance kVCT images. The selected OARs were contoured by a radiation oncologist on 18 sets of images as the ground truth. The performance of the patient-specific network was compared with the population network as well as the clinical rigid registration method. The corresponding dosimetric impacts resulting from different segmentation and registration methods were also investigated. <h3>Results</h3> The initial Dice similarity coefficient (DSC) results of the population network on the patient-specific evaluation set were 0.821, 0.537, 0.739 for esophagus, larynx, and pharynx respectively, and they were markedly improved to 0.887, 0.889, 0.850 with the proposed patient-specific network. These results also outperformed the results of 0.670, 0.792, 0.709 with the registration-based method in current clinical practice. The contouring accuracy of the patient-specific network gradually increased with the increase of longitudinal training cases and approached saturation with more than six training cases. The qualification result also shows the population network was adapted to the specific manual contouring style after the patient-specific learning. Discrepancies of the mean doses by using the manual contour and the patient-specific auto-segmentation result were -1.29%, -1.05%, -0.13% for the three OARs respectively, better than the differences of 4.80%, -6.56%, -1.84% using the registration contour. The OAR dose underestimation in some treatment fractions based on the registration contour was markedly improved when using the patient-specific auto-segmentation result. <h3>Conclusion</h3> Auto-segmentation of daily kVCT images was studied based on patient-specific learning to facilitate future adaptive radiotherapy. Compared with the common population network and the clinical registration method, the patient-specific network could achieve a much better contouring and dosimetric accuracy and thus is promising to become a powerful tool in adaptive radiotherapy applications.