【作者】 于辉；

【导师】 刘天歌； 赵谞博；

【作者基本信息】 燕山大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 癌症作为一种多发性恶性疾病,具有极高的致死性且难以进行手术根除,放射治疗作为一种综合性治疗方法,是治疗癌症最常用的一种手段。在医生制定放疗计划时,要注意危及器官的勾画,从而避免辐射线对健康器官造成损伤。为了减轻放射医生的工作负担,对危及器官的精准自动分割成为一项较为迫切的任务。但是危及器官的分割容易受到器官体积间的不平衡以及金属伪影等因素的影响,并且干扰因素只存在于CT图像的局部区域。因此,采用单一的分割模型对整幅CT图像中所有的危及器官进行分割会出现分割结果差距较大的问题。首先,考虑到现有的分割算法全部聚焦于整幅图像,并没有考虑图像局部信息之间的差异性,因此本文提出一种基于CT图像中单个危及器官分割难易程度进行量化的质量评价方法,旨在为后续的危及器官自动勾勒提供一个具有可参照性的量化指标,基于此指标能够在器官分割过程中为每个器官选择合适的分割模型,同时也可以提醒医生在纠正标注数据时要对更具挑战性的器官进行精确修改勾画。其次,为了预测CT图像中代表单个危及器官的分割难易程度的质量分数,本文提出了联合多分割模型进行反馈的打分机制,基于不同分割模型的器官分割结果之间具有的强相关性,生成单个器官的质量分数标签。并且,为了能够自动地评价器官质量,本文提取了图像中器官区域的相关特征,训练并预测器官的质量分数。最后,为了验证本文所提出方法的有效性,将方法在头颈部数据集和腹部数据集分别进行了实验验证。质量评价算法的预测结果与真实值之间保持了很强的相关性。为了进一步验证所提算法思想的有效性,本文将分割结果与未进行质量评价的传统分割模型的分割结果进行比较,实验结果证明对于头颈部九种小器官,整体Dice系数提高2.16%,FNR降低3.85%;对于腹部三种大器官,整体Dice系数提高0.34%,欠分割率降低3.14%。因此可以证明质量评估对分割效果有提升作用。更多还原

【Abstract】 As a multiple malignant disease,cancer is highly lethal and difficult to eradicate surgically.As a comprehensive treatment method,radiotherapy is the most commonly used method for treating cancer.When doctors make radiotherapy plans,they should pay attention to the delineation of organs that are dangerous to avoid radiation damage to healthy organs.In order to reduce the workload of radiologists,accurate automatic segmentation of endangered organs has become an urgent task.However,the segmentation of endangered organs is easily affected by factors such as imbalance between organ volumes and metal artifacts,and interference factors only exist in local areas of the CT image.Therefore,using a single segmentation model to segment all the organs at risk in the entire CT image can lead to a large gap in segmentation results.Firstly,considering that existing segmentation algorithms focus solely on the entire image and do not consider the differences between local information in the image,this paper proposes a quality evaluation method based on the difficulty of segmentation of a single threatened organ in a CT image,which aims to provide a referential quantitative indicator for subsequent automatic delineation of threatened organs,Based on this indicator,it is possible to select an appropriate segmentation model for each organ during the organ segmentation process,and it can also remind doctors to accurately modify and delineate more challenging organs when correcting labeled data.Secondly,in order to predict the quality score representing the difficulty of segmentation of a single organ at risk in a CT image,this paper proposes a feedback scoring mechanism that combines multiple segmentation models.Based on the strong correlation between the segmentation results of different segmentation models,a quality score label for a single organ is generated.Moreover,in order to achieve automatic organ quality evaluation,this paper extracts relevant features of the organ regions in the image,trains and predicts the organ quality score.Finally,in order to verify the effectiveness of the proposed method in this article,experimental validation was conducted on the head and neck dataset and the abdominal dataset,respectively.Firstly,the quality evaluation algorithm is validated,and the features within the organ region are extracted for regression training and prediction.The experimental results show that the proposed quality evaluation algorithm can automatically and accurately predict the difficulty of organ segmentation,and the predicted results maintain a strong correlation with the true values.In order to further validate the effectiveness of the proposed algorithm,this paper compares the segmentation results with those of traditional segmentation models without quality evaluation.The experimental results show that for the nine types of small organs in the head and neck,the overall Dice coefficient increases by 2.16% and the False Negative Rate decreases by3.85%;For the three major organs of the abdomen,although the overall Dice coefficient only increased by 0.34%,there was a significant decrease in the False Negative Rate,which decreased by 3.14%.Therefore,it can be proven that quality evaluation has an improvement effect on segmentation performance.更多还原