Evaluating a Personalized Deep-Learning-Based Auto-Segmentation Method for CBCT-Based Adaptive Radiotherapy

Abstract

<h3>Purpose/Objective(s)</h3> Online adaptive radiotherapy (ART) requires accurate and efficient auto-segmentation of target volumes and organs-at-risk (OARs) in mostly cone-beam computed tomography (CBCT) images. Propagating expert-drawn contours from the pre-treatment planning CT (pCT) through traditional or deep learning (DL) based deformable image registration (DIR) can achieve improved results in many situations. Typical DL-based DIR models are population based, that is, trained with a dataset for a population of patients, so they tend to average their result to the data distribution rather than individualizing to a single patient. We previously proposed a method, test-time optimization (TTO), that can individualize DL-based DIR model to each patient to improve DIR accuracy. In this work, we evaluated the accuracy and efficiency of auto-segmentation results from our proposed DIR method. <h3>Materials/Methods</h3> We used 39 patients with head and neck squamous cell carcinoma for evaluation. Each patient has a pCT with manual contours and two fractions of CBCT with manual contours. 17 structures that are either critical OARs or have large anatomical changes during radiotherapy courses were selected. These structures were: left brachial plexus, right brachial plexus, brainstem, oral cavity, constrictor, esophagus, nodal gross tumor volume, larynx, mandible, left masseter, right masseter, posterior arytenoid-cricoid space, left parotid gland, right parotid gland, left submandibular gland, right submandibular gland, and spinal cord. Dice similarity coefficient (DSC) and 95% Hausdorff Distance (HD95) were calculated between auto-segmentation and manual segmentation for each structure. All the auto-segmentation methods tested in this experiment are based on DIR methods including our proposed individualized DIR models and traditional DIR models. To get individualized DIR model for each patient, TTO was applied where a pre-trained population DIR model was fine-tuned on each patient separately. The auto-segmentation accuracy and the time cost for DIR model individualization were studied. <h3>Results</h3> The average DSC and HD95 of the auto-segmentation by individualized DIR models is 0.85 and 2.33mm respectively over 17 selected OARs and a target of 39 patients, which performs better than traditional DIR methods. The average time for deriving the individualized DIR model is approximately 3.6 minutes. <h3>Conclusion</h3> The proposed TTO method is well-suited for online ART and can boost segmentation accuracy for DL-based DIR models, especially for outlier patients where traditional DIR methods fail.