| One key procedure of accurate radiation therapy is to accurately and efficiently delineate organs at risk(OARs)and targets on medical images,which affects the accuracy of radiotherapy treatment and treatment outcomes.At present,OAR auto-delineation algorithms can be categorized into medical image registration and medical image segmentation,which usually suffer from these challenges: 1)trade-off between performance and computational time in 4D image registration.Pairwise registration ignores the temporal coherence across motion phases(e.g.,respiration),thereby generating inferior deformation vector fields.In contrast,groupwise registration incurs high computational costs and long optimization time.Besides,deep learning based registration requires considerable labeled data for training and exhibits inferior performances than optimization-based methods.2)the problem of few training data.Annotating medical images requires clinical medical expertise and is extremely time-consuming,which makes it difficult to collect large medical datasets,especially for segmentation tasks that require voxel-wise labels.In deep learning,training with small datasets will cause model overfitting to the training data and degenerate the segmentation performances.3)the quality of CBCT scans.Compared to other medical modalities like magnetic resonance imaging,CBCT scans have lots of artifacts and noises,low contrast between soft tissues and low image quality,which causes that the model cannot precisely distinguish the vague organ boundary to achieve auto-delineation.To solve the above challenges,we propose a deep learning based few-shot groupwise registration algorithm to solve the optimization problem in registering 4DCT scans by using a few-shot registration network.The few-shot registration method exploits an implicit reference image to mitigate the bias introduced by the selection of reference images.The pre-training with few labeled images facilitates the acceleration of the optimization process.As for the problem of few training data,a registration based semi-supervised segmentation algorithm is presented.The proposed method propagates manual delineation of unlabeled images onto abundant unlabeled images by utilizing deformable image registration,and trains the head and neck organ segmentation model with the generated pseudo-labels,which mitigates the overfitting and inferior performance of neural networks trained with few images.To overcome the segmentation of CBCT images with low quality,we devise a deformable image registration guided segmentation framework.The proposed method generates initial contours for CBCT scans with Demons registration methods and helps the model focus on extracting the local features around the initial contours to achieve accurate final segmentation.To verify the effectiveness of the proposed algorithms,comprehensive experiments are conducted on multiple datasets of 4DCT,CT and CBCT scans.The experimental results show that our methods solve the above challenges and improve the model performance on registration of 4DCT and segmentation of CT and CBCT scans,which demonstrates its prospective applications in OAR delineation in adaptive radiotherapy. |
