中国科学院分子影像重点实验室知识库

       临床上对疾病的精确诊断能够帮助医生更好地开展临床治疗，改善患者的预后。而临床上对疾病的准确诊断主要依赖于穿刺活检或者术中病理等有创的方法，但这类方法风险、成本高而且伴随并发症，难以广泛推广。因此，临床上亟待发展无创安全、准确的诊断方法。

       随着医疗成像技术和成像设备的发展，医疗影像在疾病诊断中起到了重要的作用。其中超声成像凭借价格低廉和安全方便等特点，在乳腺等器官的检查和疾病诊断上得到广泛应用。但超声图像同时也存在分辨率相对较低、成像效果严重依赖于操作者等缺点，难以被定量分析，导致对超声图像的分析主要依赖于医生的临床经验，而难以得到高水平的广泛使用。

       近几年，随着人工智能技术的发展，基于人工智能的影像组学分析方法能够实现医疗图像的定量分析，而其中基于深度学习的分析方法能够克服超声图像分辨率较低、轮廓边缘模糊等缺点，能够自主地从大量的超声图像数据中挖掘与诊断结果相关的高纬度特征，实现疾病的精确诊断。目前超声影像组学发展还处于早期阶段，需要做大量的探索和尝试才能更好地发挥超声图像的临床应用价值，为患者提供准确、方便和成本低廉的疾病诊断方式。因此，本文将利用深度学习，对常规超声图像和超声弹性图像两种模态的超声图像进行定量分析，着眼于可用超声图像诊断的相关疾病，探索能够利用超声图像准确诊断疾病的模型方法。总的来说，本文开展了以下两个方面研究。

       针对术前肿块经核心针活检（core needle biopsy，CNB）诊断为纯乳腺导管原位癌（ductal carcinoma in situ, DCIS）的患者的术后病理可能发生升级而导致需要进行二次手术的问题，本文以CNB诊断为纯DCIS的患者术后升级预测为研究内容，研究如何搭建卷积神经网络模型分析乳腺常规超声图像，实现对DCIS患者术后升级的准确预测。模型以ResNet神经网络模型作为基础，并根据乳腺肿块范围变化较大等临床问题做出适当的结构设计，最终模型在训练集和验证集上受试者工作特征曲线面积（Area Under Reciever Operating Characteristic Curve, AUC）达到了0.822和0.802，在同类研究中处于较好的水平，且通过三折交叉验证证明了模型具备较好的鲁棒性。本研究展现了通过基于深度学习的超声影像组学对DCIS患者的术后升级预测的潜力。

        针对超声弹性图像测量值对于不同种病因的肝纤维化分期存在诊断阈值差异，难以统一标准的临床问题，本研究以多种病因肝硬化诊断为研究对象，研究如何设计卷积神经网络模型分析超声弹性图像，实现多病因肝纤维化患者的肝硬化诊断，解决弹性测量值在肝纤维化分期诊断阈值难以统一的问题。本文基于ResNet网络结构，利用超声弹性图像测值量程对特征层进行缩放实现图像特征的标准化，使得模型能够分析不同参数设置的超声弹性图像，同时对超声弹性图像做图像格式的转化和特征序列拼接2D-SWE测量值来提高模型的鲁棒性和诊断性能，最后结合支持向量机实现更高精度的分类诊断。模型在训练集和验证集上对肝硬化的诊断结果略优于临床统计学分析方法，在多个以病因和机器划分的测试集上的结果表明模型具备较好的鲁棒性和泛化性，展现了利用深度学习模型和超声弹性成像实现多病因肝纤维化患者肝纤维化分期的高可行性。

  Accurate diagnosis of diseases can help doctors carry out a more appropriate clinical treatment, improving the prognosis of patients. The clinically accurate diagnosis of diseases mainly relies on needle biopsy or intraoperative pathology. However, the risk and cost of these invasive methods are high and they are accompanied by complications that make it difficult to be widely promoted. Therefore, non-invasive, safe, and accurate diagnosis methods in the clinic are urgent to develop.

  With the rapid development of medical imaging technology and imaging equipment, medical imaging has become an important method of disease diagnosis.

  Ultrasound is cheap, safe and convenient, and is widely used in disease diagnosis, especially in superficial organ examinations. However, ultrasound images have the disadvantages of low resolution and their quality is heavily dependent on the operator, making it difficult to quantify. The accurate analysis of ultrasound images mainly depends on the doctor's clinical experience, which makes it difficult to be widely used at a high level.

  In recent years, with the development of artificial intelligence, the image analysis method based on artificial intelligence can realize the quantitative analysis of medical images, and the analysis method based on deep learning can overcome the shortcomings including low resolution and blurred contour edges of ultrasound images. Deep learning can autonomously mine the high-latitude features related to the diagnosis results from a large number of ultrasound images, and realize the accurate diagnosis of the disease. At present, the development of Radiomics based on ultrasound images is still in its infancy, and a lot of exploration and experimentation is needed to tap more clinical application value, providing patients with accurate, convenient, and low-cost diagnoses. Therefore, we will use Radiomics based on deep learning to quantitatively analyze the two modalities of ultrasound images, including conventional ultrasound images and ultrasound elastography. This paper focuses on the relevant diseases that can be diagnosed by ultrasound images and exploring more accurate diagnosis methods. In general, this paper will carry out the researches from the following aspects.

  The postoperative pathology of patients who were diagnosed as pure ductal carcinoma in situ (DCIS) by core needle biopsy (CNB) before surgery may upstage to micro-infiltration, which would lead to the need for a second surgery. This paper focuses on predicting postoperative upgrade of pure DCIS diagnosed by CNB before surgery and uses a convolutional neural network to analyze conventional breast ultrasound images to achieve accurate prediction of postoperative upgrade of DCIS patients. The model of neural network derives from ResNet, and we make some appropriate adjustments according to clinical problems such as the great changes in the size of masses. On the training set and validation set, the Area Under Reciever Operating Characteristic Curve (AUC) reached 0.822 and 0.802, which is higher than other similar studies. Three-fold cross-validation has been done to verify the great robustness of the model. This research has proved the potentiality of ultrasound Radiomics for postoperative prediction of DCIS patients.

  Aim at solving the clinical problem that ultrasound elastography has different value standard for liver fibrosis of different causes, this paper focuses on liver cirrhosis that causes by various disease and use a convolutional neural network to accurately predict liver cirrhosis. In our deep learning network bases on ResNet, we zoom the feature sequence with the measurement range of ultrasound elastography to achieve image standardization, making it possible to make use of the ultrasound elastography with different parameter settings. The conversion of image format, concatenation of ultrasound elastography value and feature sequence are done to improve the robustness and diagnostic performance of the model. We make use of a support vector machine to achieve a more precise diagnosis. Finally, the diagnosis of liver cirrhosis on the training set and validation set is slightly better than the traditional clinical statistical analysis. The great robustness and generalization of the model are verified through multiple test sets that are divided by pathogens and machines. The research demonstrates the high feasibility of using deep learning models and ultrasound elastography to achieve liver fibrosis staging in patients with multi-cause liver fibrosis.