【作者】 魏涛；

【导师】 赵亮；

【作者基本信息】 广西大学 ， 计算机科学与技术， 2024， 硕士

【摘要】 肝脏是人体内最重要的器官之一,对维持人体内众多生理功能的正常运行有着重要意义。然而肝脏相关疾病却具有发病率高、致死率高的特点,严重威胁着人民的生命健康安全。相关研究发现,早期肝癌患者的五年存活率远远高于晚期肝癌患者,因此对于肝脏疾病的早期筛查十分重要。计算机断层成像技术是对肝脏疾病进行检查时最常见的技术,该技术与近年来蓬勃发展的深度学习技术相结合,已经实现了高效、精准的肝脏病灶自动分割。但是,当前众多分割方法在面对低对比度病灶、微小病灶、非光滑模糊病灶时往往表现不佳,需要进行进一步的优化。为了解决上述问题,本文对现有肝脏病灶分割方法进行了改进,具体研究内容如下:针对当前模型面对低对比度病灶、微小病灶时分割困难的问题,本文将伪装目标检测方法引入肝脏病灶分割领域,将解码器设计为搜索-识别的二阶段模式。在搜索阶段中,模型基于骨干网络提取的多尺度特征分别生成区域引导信息和关注细节的引导信息,在识别阶段中,模型结合两种引导信息渐进、迭代地对分割预测进行精细化处理,以实现准确的区域预测。针对当前模型面对非光滑模糊病灶时分割困难的问题,本文将路径签名方法引入肝脏病灶分割领域。本文首先从引导信息中提取出骨架边界预测,然后计算其与骨架边界掩码各自的路径签名,并基于余弦相似度设计了全新的路径签名损失函数对骨架边界进行监督,最终实现了精准的边界预测。为了将两种预测有效结合,本文引入多任务学习策略,将区域预测任务嵌入区域分支,将边界预测任务嵌入边界分支,并对双分支同时进行训练,最终得到了性能优异的肝脏病灶分割模型。本文在Li TS公开数据集以及太和医院SLS数据集上进行了大量实验,结果表明本文提出的PCNet+模型明显优于其他10种先进的对比模型。这证明了伪装目标检测方法在分割低对比度病灶、微小病灶时的优越性,也体现了路径签名方法在分割非光滑模糊病灶时的有效性。更多还原

【Abstract】 The liver is one of the most important organs in the human body,and it plays a crucial role in maintaining many physiological functions.However,liver-related diseases are characterized by high incidence and mortality rates,which seriously threaten people’s life and health.Relevant studies find that the five-year survival rate of early-stage liver cancer patients is much higher than that of advanced-stage liver cancer patients,which highlights the importance of early diagnosis of liver diseases.Computed tomography is the most common technique for examining liver diseases,and its combination with the booming deep learning technology achieves efficient and accurate automatic segmentation of liver lesions.However,many current segmentation methods often perform poorly when facing low contrast lesions,small lesions and lesions with non-smooth blurry boundaries,which need to be further optimized.In order to solve the above problems,this study improves the existing liver lesion segmentation models,and the specific research content is as follows:To tackle the problem of segmenting low contrast lesions and small lesions,this study introduces the method of camouflaged object detection into the field of liver lesion segmentation,and designs the decoder as a searchidentification two-stage paradigm.In the search stage,the model generates region guidance information and details-concentrated guidance information based on multi-scale features extracted by the backbone network.In the identification stage,the model refines the segmentation prediction progressively and iteratively based on the two guidance,which achieves accurate region prediction.To solve the problem of segmenting lesions with non-smooth blurry boundaries,this research introduces the method of path signature into the field of liver lesion segmentation.In this research,the skeleton boundary prediction is extracted from the guidance information,and the path signatures of both the skeleton boundary prediction and the skeleton boundary mask are calculated.Besides,this study designs a novel path signature loss function based on cosine similarity to supervise the skeleton boundary,which achieves precise boundary prediction.In order to effectively combine the two predictions,this study employs a multi-task learning strategy,embedding the region prediction task into the region branch,embedding the boundary prediction task into the boundary branch,and training the two branches at the same time to obtain a liver lesion segmentation model with excellent performance.Extensive experiments conducted on the Li TS public dataset and the SLS dataset from the Taihe Hospital show that the proposed PCNet+ is significantly superior to other 10 advanced comparison models.This proves the superiority of the camouflaged object detection method in segmenting low contrast lesions and small lesions,and also demonstrates the effectiveness of the path signature method in segmenting non-smooth blurry lesions.更多还原