Research On Automated Multi-Organ Segmentation In Precise Radiotherapy

With the rapid development of computer technologies,oncology radiation has entered an era of precise treatment.Image-guided radiation therapy can not only deliver enough radiation dose to kill the tumor tissues,but also protect the adjacent normal structures effectively,which improves the survival rates and life quality of the patients greatly.Before radiotherapy treatment,tumor regions and organs-at-risks(OARs)need to be manually delineated by oncologists and radiotherapy physicists.With the accurate delineations,radiotherapy simulations can be conducted by planning system to obtain the ideal dose distributions.Combining the computer technologies and algorithms,automated segmentation methods can achieve the tumor and multi-organ segmentation accurately and efficiently,which may replace the manual delineation,have become the critical techniques to improve the radiotherapy efficiency and performance.However,different imaging modalities and different anatomical structures bring great difficulties to accurate segmentation of OARs and tumor tissues.Owing to the high representation ability and discriminability,deep learningbased technologies have achieved promising results in various image processing tasks.This dissertation aims to tackle with the challenges in multi-organ segmentation in various anatomical environments for patients undergoing radiotherapy treatments,improve the segmentation efficiency and performance of multi-organs,and provide potential automatic segmentation tools for clinics.The main contributions of the dissertation can be summarized as follows:1.In automated segmentation of multi-organ with large differences in organ shapes and sizes,precise detecting the complete shape of multi-organ is highly difficulty,especially in computed tomography(CT)image with low soft tissue contrast.Moreover,there are a lot of soft OARs in head and neck regions,such as brainstem,parotid glands,submandibular glands,larynx,pharynx,and so on,slight differences of organ delineations may result in serious side effects in radiotherapy treatment effect.Thus,keeping accurate and complete organ shapes in segmentation results is of great significance.To achieve accurate and efficient segmentation of multiply head and neck organs,a novel shape representation model(SRM)is integrated in the proposed segmentation framework to constrain the shapes in the predicted segmentation results.The proposed SRM is pre-trained on the binary masks of the ground truths in the training set to learn the shape characteristics of the organs to be segmented.The segmentation network is a U shape network,which is composed of residual blocks.During training,the pre-trained SRM is utilized to project the segmentation results and corresponding ground truths into the latent shape space to obtain its latent shape representations.The differences between them are then incorporated into the loss function to improve the shape quality of the segmentation results.Experimental results on the public head and neck CT dataset demonstrate the segmentation ability and effectiveness of SRM.2.Although SRM shows high effectiveness in reducing the difference between the network predictions and ground truths in latent shape space,segmentation network achieves image-level voxel-wise classification,global contexts of prediction are ignored and need to be constrained.Aiming to deal with the difficulties in automated head and neck multiorgan segmentation in low-field magnetic resonance(MR)-guided radiation therapy system,a segmentation framework which combines dense connected network,adversarial training mechanism,and SRM is proposed.In the proposed framework,SRM and convolutional neural network(CNN)based discriminator are employed to constrain the network predictions in latent shape space and image-level,respectively,and thus to strengthen the network performance.The effectiveness of the proposed framework has been validated on the public head and neck CT and low field MR data sets,respectively.3.The size of medical image dataset with detailed manual delineations are commonly limited,which limits the extensive training of complex 3D segmentation network and easily results in over-fitting.Moreover,the complete ground truths are more difficult to obtain.Attention mechanism is an idea technique to enhance the useful features that are highly-related to the targets,weaken the redundant features,and improve the network representation capability without increasing the complexity of the network greatly,which are of great significance for limited training dataset available.Thus,to boost the network performance with limited training dataset,channel-wise and spatial attention modules are incorporated to the sub blocks to model and recalibrate the relationships among the features.The performance of the enhanced segmentation network has been evaluated on the public dataset.Moreover,the experimental results on a weekly head and neck dataset demonstrate that the network is able to detect the anatomical changes of the OARs during radiotherapy treatment.4.Towards the challenges in abdominal multi-organ segmentation,including large variations in segmentation difficulties,blurry boundaries,and irregular organ shapes,a task-wise self-paced Dense Net is proposed.During training,the loss terms for individual organs are weighted based on the learning paces of individual targets to balance the network optimization process.At the same time,boundary constrain loss term are introduced in the loss function to enhance the edge detection ability and improve the integrity and consistency of organ boundaries.A public abdominal CT dataset with manual delineations of eight organs was employed to verify the performance of the proposed task-wise self-paced Dense Net.The experimental results confirm the advantages of the task-wise self-paced learning mechanism and boundary constraint loss term.5.To apply the proposed head and neck multi-organ segmentation network to clinical practice,an automated segmentation plug-in was developed and integrated into the software system in radiotherapy department in hospital.The segmentation plug-in achieves the integration of delivering scanning data,automatic segmentation,and visualization of the segmentation results,which improves the workflow efficiency greatly.Moreover,the segmentation performance and reliability of the plug-in has been highly praised and affirmed by doctors and physicists.