Purpose: In cone-beam CT (CBCT)-guided adaptive radiotherapy (ART), deep learning (DL)-based automatic organ-at-risk (OAR) delineation can be an important time saving that enables an online ART workflow. However, there remain several barriers to the effective implementation of these approaches in a clinical setting, including severe image artifacts in CBCT images that hinder the delineation of critical structures and the difficulty of curating a training data set consisting of consistent manual annotations. We propose here a patient-specific auto-contouring approach utilizing the Intentional Deep Overfit Learning (IDOL) framework that demonstrates a significant improvement in contouring accuracy using only a few training data sets.
Methods: IDOL-based auto-contouring is an application of intentional overfitting that leverages patient-specific prior information. Training in the IDOL framework involves two stages: 1) conventional training of a generalized segmentation model with a training set of 30 CBCTs and manual contours, and 2) using the planning CT and contours of a 31st patient as prior information, the generalized model is trained further to achieve the IDOL model that is applied to the CBCT of the 31st patient. For testing, step 2 was repeated for patients 31-39 using the same generalized model as a starting point.
Results: The proposed method was evaluated on a cohort of 9 head and neck cancer patients. The generalized model (step 1 training only) achieved Dice similarity coefficients (DSC) of 0.80 and 0.64 for the larynx and esophagus, respectively. By adopting the personalized IDOL model in each test case, the DSC improved to 0.89 (+0.09) and 0.74 (+0.10) for the same structures, showing a significant improvement in contouring accuracy.
Conclusion: In this study, we have successfully applied the IDOL framework to the CBCT auto-contouring task. Our approach is widely applicable to common tasks in radiotherapy and can be utilized for CBCT or MRI-guided ART.
Not Applicable / None Entered.