神经网络论文 - Translated by Marianne Wang
Thesis on Neurology

颅内胶质瘤尤其是恶性胶质瘤术后早期行影像学检查的主要目的是评价手术效果，发现残存肿瘤，以便制定进一步的治疗方案，但是肿瘤残存与正常脑组织的反应性增强在影像学上很难鉴别。目前多数学者认为磁共振(MR)检查是鉴别肿瘤残存与正常反应的首选检查方法[1，2，3]，并提出了鉴别诊断的特征或标准[1，2]。但是，肿瘤残存与正常反应在影像特征上往往有重叠，另外医师的诊断水平容易受主客观因素的影响，因此做出准确的鉴别诊断有时仍是困难的。我们利用具有人工智能特性的人工神经网络(artificial neural network, ANN)进行辅助鉴别诊断，取得了满意的效果，现报告如下。
The purpose of early image examination of intracranial gliomas, especially malignant intracranial gliomas, is to evaluate the results of the surgery, to locate tumor residual so as to determine the next treatment. However, it is very difficult to differentiate tumor residual from benign enhancement on the basis of images. Presently, many scholars hold the opinion that Magnetic Resonance (MR) is the ideal examination approach to differentiate tumor residual and benign enhancement[1, 2, 3] , and have further stated the features or standards[1, 2] for differentiation and diagnosis. However, tumor residual and benign enhancement often overlap in terms of their image characteristics. Besides, due to the subjectivity of the diagnosis abilities of doctors, it is still difficult to provide accurate differentiations and diagnoses. It is worth mentioning that we have employed Artificial Neural Network (ANN) with artificial intelligence to assist to differentiate and diagnose, which already shows satisfying effects. The results are as follow:

材料与方法
Materials and Methods

一、 临床资料
抽取1991年10月～2000年12月间我院经手术与病理证实的胶质瘤病例100例，进行回顾性分析，其中男73例，女27例，年龄12-62岁，平均35.2岁。其中星形细胞瘤45例，少突胶质细胞瘤21例，胶质母细胞瘤25例，混合型胶质瘤9例。幕上95，幕下5例。
A. Clinical Data
100 cases of intracranial gliomas confirmed by operation and pathology in our hospital from Oct. 1991 to Dec. 2000 were analyzed retrospectively. Of the 100 cases, 73 of the patients are male, 27 female, their ages ranging from 12 to 62 with an average age of 35.2. Astrocytoma 45 cases, oligodendroglioma 21 cases, colloid female cytoma 25 cases, mixed gliomas 9 cases. Above tentorium 95 cases, below tentorium 5 cases.
二、 检查方法
MR机型为美国GE公司signa 1.5T和0.5T超导机，扫描序列为自旋回波或快速自旋回波序列，矩阵256×256，层厚3～5mm。每位病人均进行矢、冠和轴位平扫及增强扫描，对比剂为北京北陆公司生产的磁显葡胺，静脉注射剂量为0.1mmol/kg。所有病例术前均进行MR检查，并在术后一个月内至少进行一次复查，最早在术后24小时之内。以后定期复查，直至软化灶形成或肿瘤复发。
B. Examination Method
The MR machine are Signa 1.5T and 0.5T superconductors produced by GE. The scanning sequences are spinning echo wave or rapid spinning echo wave, matrix 256x256, thickness 3~5mm. Each patient has been scanned sagittal, coronal, axial, and enhanced. The contrast-medium is 磁显葡胺 produced by Beijing Beilu Co. with the dose of intravenous injection as 0.1mmol/kg. All of the cases were examined by MR before the surgery, and were reexamined at least one time within one month after they surgery, the earliest being within 24 hours after the surgery. Then regular follow-ups by MR were conducted until the malacia appeared or the residual tumor grew again.

三、 磁共振诊断标准
1.肿瘤残存
术后早期行MR检查即出现的增强，追踪观察并无减弱或消失的趋势，直至增强部位有肿瘤的复发。术后残存与术后反应性增强同时存在时，视为肿瘤残存。
2.肿瘤复发
出现以下任一情况考虑肿瘤复发：①术后追踪观察，在术区周围或远处出现新的增强灶：②原有残存增强的病灶增大20%以上；③如果原有残存增强的病灶增大小于20%，但是出现以下任一情况高度怀疑复发：a 水肿或占位征象重新出现或加重：b 临床症状加重。
3.术后反应性增强
术后早期在术区边缘出现的增强，追踪观察有逐渐减弱或消失的趋势，直至在原增强部位形成软化灶。术前未出现增强的肿瘤在术后早期(一周之内)出现的增强认为是术后反应性增强。

C. MR Diagnosis Criteria
1. Tumor Residual
Enhancement that emerged right after the surgery during MR examination showed no tendency of reduction or disappearance until the reappearance of tumor at the enhanced part. When the postoperative residual and postoperative reactional enhancement coexist, it is regarded as tumor residual.

2. Tumor Recurrence
Tumor recurrence is considered when: ① new enhancement appears around or away from the operative area, ② the original residual focus increases by 20% or above; ③ the original residual focus increases by less than 20%, but either of the following conditions highly suggests recurrence: edema or occupancy reappears or becomes worse, b. clinical symptoms deteriorate.

3. Postoperative Reactional Enhancement
Enhancement emerged around the operative area in early postoperative period shows tendency of gradual reduction or disappearance until malacia appears in the original enhancement area. Enhancement of tumor emerged during early postoperative period (within one week) that has not shown up before surgery is regarded as postoperative reactional enhancement.

四、 人工神经网络
我们使用的网络是美国加利福尼亚科学软件公司的商业软件包(BrainMaker Professional, version 3.74)，运行计算机中央处理器：英特尔赛杨466MHz，操作系统：微软公司视窗98，网络算法：误差反向传播算法(back propagation, BP算法)，使用的传递函数是sigmoid函数。训练容许误差定为0.1。网络拓扑结构分3层：输入层、隐含层、输出层，输入节点数取5，隐含节点数取6，输出节点数取1。网络拓扑结构见图1。
在仔细分析描述术后的MR图像之后，提取有鉴别诊断意义的五个指标作为输入数据，这五个指标是增强与术前是否有对应关系、环形增强厚度是否均匀、环形增强厚度大小、边界是否清楚、是否伴结节或团块状增强，见表1。规定肿瘤残存时输出节点值为1，反应性强化输出节点值为0。一个病例的输入输出数据组成一个样本，多个样本组成数据组。将所有数据分为训练数据、检验数据，在网络训练过程中，不断用检验数据来检查网络的训练情况，以防止网络过度训练。另外保留一组新的数据共20个病例作为检验数据对训练后的网络进行检验，这组检验数据又叫做产品数据。
D. Artificial Neural Network
We have employed the network of BrainMaker Professional, version 3.74 provided by California Scientific Software Inc. CPU of operating computer: Intel Celeron 466MHz, operating system: Microsoft Windows98, network algorithm: Back Propagation (BP), transfer function: sigmoid function. The allowed practical corrigendum is 0.1. The network topology is in three layers: input layer, hidden layer, output layer. The point of input layer is 5, the point of hidden layer is 6, and the point of output layer is 1. Please refer to Diagram 1 for the network topology.

After a careful study and detailed description of postoperative MR images, 5 indexes meaningful for differentiation and diagnosis are extracted and regarded as input data. These five indexes are: correlation between enhancement and preoperative conditions, evenness of annular enhancement, thickness of annular enhancement, clearness of margin, existence of nodular or ball enhancement, please refer to table 1. The value of output point for tumor residual is fixed as 1, the value of output point for reactional enhancement is fixed as 0. The input and output data of a case comprises a sample, and multiple samples comprise a data group. All data are categorized into training data and examination data. During the course of network training, examination data are constantly involved to examine it so as to prevent overtraining. Another data group consisting of 20 cases is regarded as examination data to examine the network after the training. These data group is also called product data.

五、网络性能评价
作为产品数据(检验数据)的病例(共20例)经一位有经验高年资放射学医师在不知道随访检查结果和网络诊断结果的情况下阅片，医师的诊断水平与网络诊断性能用受试者操作特征(receiver operating characteristic, ROC)方法进行比较。用Spss10.0软件进行ROC分析及曲线绘制。
E. Network Performance Evaluation
The cases (altogether 20 cases) regarded as product data (examination data) were reviewed by an experienced senior radiologist who did not know follow-up examination results and network diagnosis results. The diagnosis ability of the radiologist and the network performance were compared by Receiver Operating Characteristics (ROC). The ROC analysis and the curve diagram were made by Spss10.0.
结 果

用训练数据对网络进行训练，最后迭代次数为200次左右网络诊断误差达到容许水平(0.1)。用检验数据对网络进行检验，结果表明当输出阈值定在0.9时，网络诊断出13例残存肿瘤的10例，敏感性是76.9%(10/13)，诊断出7例反应性强化中的6例，特异性是85.7%(6/7)，见图2。放射学医师诊断出8例残存肿瘤，敏感性是61.5%，诊断出5例反应性强化，特异性是71.4%，神经网络ROC曲线下面积Az是0.813±0.105，显著高于放射学医师ROC曲线下面积Az= 0.665±0.130。ROC曲线见图3。
Results

The training data were employed to train the network. When iterations reached around 200 the error of network diagnosis would fall within the acceptable level (0.1). Examination data were employed to examine the network, and the results showed that when the threshold output value was 0.9 the network could diagnose 10 cases out of the 13 tumor residual cases with a sensitivity of 76.9%(10/13), and 6 cases out of the 7 reactional enhancement cases with a specificity of 85.7%(6/7), please refer to Diagram 2. The radiologist diagnosed 8 tumor residual cases with a sensitivity of 61.5% and 5 reactional enhancement cases with a specificity of 71.4%. The neural network ROC Az is 0.813±0.105, which is much higher than the radiologist ROC Az=0.665±0.130. Please refer to Diagram 3 for ROC curve.
讨论

颅内恶性胶质瘤是中枢神经系统常见肿瘤，具有无限增殖和浸润生长的特征，目前的外科手术难以做到根治性切除，术后及早给予放疗、化疗等综合治疗，可以延长病人的生存时间。外科手术作为一种创伤性治疗方法必然引起术区周边发生一系列病理改变，术后早期行MR检查会在术区周边出现反应性环形增强，这种反应性增强与肿瘤残存引起的增强表现相似，有时很难鉴别。国内外众多的学者对术后早期反应性强化与肿瘤残存的鉴别进行了系统的研究[4，5]。张红梅等[1]经过大量病例的回顾性分析提出了对二者的鉴别具有显著性差异的判别指标，它们是：增强与术前是否有对应关系、环形增强厚度是否均匀、边界是否清楚、环形增强厚度大小、是否伴结节或团块状增强。以这些指标作为判别依据使得放射科医师对二者的鉴别诊断水平达到了一个新的高度。
很明显以这五项指标做判断仍然是一种多因素的线性判别模式，反应性强化与肿瘤残存影像特征复杂而且往往有重叠，不能单纯由某一项指标对某一病例是反应性增强或是残存增强做出较为肯定的判断。做出准确的诊断要综合考虑这几种指标，并易受各种主客观因素的影响。人工神经网络是一种非线性判别模式，非常适合这种多变量的模式识别，优于传统的线性判别分析[6，7]。
人工神经网络是模拟生物大脑神经元结构和功能，由彼此之间高度连接的处理单元构成的一种信息处理系统[8，9]。神经网络通常由三层单元(节点)或人工神经元组成。第一层即输入层，包含一组处理单元，负责接受向网络输入的数据。输入层与隐含层相连接，隐含层既不直接接受输入数据也不直接产生输出数据。不同的网络可能含有一层或多层隐含层，最后一层与输出层相连接。每一个输入单元都与隐含层的隐单元相连接。隐含层最后一层的隐单元都与输出层的单元相连接。输出层含有一个或多个节点, 这些节点产生输出数据。
神经网络的不同节点之间的连接权重是不同的，每个连接之间有不同的权重，并在训练期间发生变化。网络训练的过程是用一组输入数据与相应的输出数据输进网络，网络根据这些数据来调节不同连接节点之间的权重，直至实际输出与期望输出相同或相似。通过输入大量数据可以使各个训练组的网络误差减小。这种改变节点间连接权重的方法类似于大脑的突触对刺激信号进行调节的现象：突触受到刺激以后，刺激的权重之和传送给细胞核，权重和大于这个神经细胞的阈值，神经细胞便被激发，并给出输出。受到刺激信号的神经细胞的输出只有两种情况，激发或未被激发。大脑的学习过程就是神经元之间连接强度随外部刺激信息做自适应变化的过程，大脑处理信息的结果确由神经元的状态表现出来。
人工神经网络在放射学中有着广泛的应用，但是国内应用报道较少[10]。神经网络是一个高度非线性化系统，具有并行结构和并行处理、知识的分别存储、高度的容错性、自适应性和联想力，它能够学习和记忆输入模式的特征，并将该特征应用于新个体的诊断[11]。本研究所用的BP网络是迄今为止用得比较广泛和流行的最普通的网络[11]。它的主要思想是把学习过程分为两个阶段：第一阶段(正向传播过程)，给出输入信息通过输入层经隐含层逐层处理并计算每个单元的实际输出值；第二阶段(反向过程)，若在输出层未能得到期望的输出值，则逐层递归地计算实际输出与期望输出之差值(即误差)，以便据此调节权值。
要使训练出的网络达到最佳性能并能够实际应用，这就涉及到网络的结构和设计问题。在网络设计的开始阶段要确定输入数据，输入节点能够代表每个数据，所以要弄清楚正确的数据源，如果数据源中包含大量虚假或不相关的数据，那必将妨碍对网络的正确训练[12，13]。在本研究中，我们提取对正常反应性强化与肿瘤残存有鉴别意义的5个磁共振图像特征作为输入数据输入网络，而舍弃了相关性不大的临床及影像特征，如肿瘤病理类型、复查时间、水肿情况、环形增强是否完整等。据文献报道，胶质瘤术后反应性强化的发生率是65.7%(134/204)，肿瘤残存发生率是34.3%(70/204)[1]，因为网络具有学习和记忆功能，能够记住数据的分布特点，因此用此比率的数据训练网络，对反应性强化会过度学习，即使网络训练不成功，全部诊断为反应性强化，正确率仍然高达65.7%。因此，本研究所用数据反应性强化与肿瘤残存病例各约占50%，使网络能“公正”地学习二者的磁共振影像特点。另外，基于BP算法的神经元网络中隐含层节点数的选择是影像网络性能的一个关键因素，所以，要选择恰当的隐含层节点数，但是，目前仍没有一个好的方法确定隐含层节点的数目，往往凭经验或通过试验来确定[12，13]。隐含层数目经过多次试验，最终定为6，网络达到了较高的诊断水平。
本研究显示人工神经网络对新个体的诊断特异性是85.7%，敏感性是76.9%，为评价网络的诊断水平，我们用ROC的方法对其与高年资放射学医师的诊断水平进行比较，ROC曲线下面积是Az=0.813±0.105，高于放射学医师的诊断水平(Az=0.665±0.130)。ROC方法已经成国际上公认的实际应用最为广泛的评价2种或2种以上影像诊断系统的标准方法[14]。神经网络的正确训练与使用有赖于有经验的医师对术后磁共振图像强化的正确描述，应采用一致的标准，我们所用的5个指标在实际操作当中简单易行，可以保证网络的训练者与使用者对图像描述的一致性，发挥网络的最高性能。
本研究成功地证明了人工神经网络能够鉴别颅内恶性胶质瘤术后磁共振图像的反应性强化与肿瘤残存所致的强化，为解决二者的鉴别这一复杂问题提供了一个新的方法，但是本研究只是初步的应用探讨，能否在实际中应用，还有待于进一步的研究，如训练样本的增加、图像描述的规范化、对医师诊断影响程度、参数的优化等，网络的性能会不断完善和提高。随着医师对人工神经网络技术的熟悉和认可，它完全有可能一种实用的工具，辅助放射科医师做出更准确的诊断。
Discussion

Intracranial gliomas is common for central nervous system and is featured for its unlimited reproduction and infiltration. Presently it is difficult to radically remove it with a surgery. Comprehensive treatment including radiotherapy and chemotherapy right after the surgery will prolong the lives of the patients. As a traumatic treatment, a surgery will for sure cause a series of pathologic changes around the operated area. At the early postoperative period, MR examination will cause reactional annular enhancement around the operated area, which is very similar with the enhancement caused by tumor residual. Sometimes it is difficult to differentiated them. Many scholars home and abroad have carried out systematic researches[4, 5] regarding the differentiation between early postoperative reactional enhancement and tumor residual. Zhang Hongmei[1], together with other scholars, proposed via many retroactive analyses the determining indexes that greatly differentiate the two, viz.: correlation with the enhancement and preoperative conditions, evenness of the annular enhancement, clearness of margins, thickness of annular enhancement, existence of nodular or ball enhancement. As determining references, these indexes greatly improve the diagnosis ability of a radiologist.

It is apparent that determination from these five indexes is still a linear determination mode. The characteristics of reactional enhancement and tumor residual images often overlap, and no positive determination whether a case belongs to reactional enhancement or tumor residual can be made by a single index. To make an accurate determination, these indexes shall be considered comprehensively with an effort to eliminate any subjective influences. The artificial neural network is a non-linear determination mode, which is more optimal than the traditional linear determination analysis[6, 7] and is ideal for the identification of the mode with such multiple variables.

Artificial neural network simulates the architecture and functions of a biological cerebral neuron. It is a information processing system[8, 9] constituted by processing units tightly connected with each other. A neural network is usually comprised by three layers of units (nodal points) or artificial neurons. The first layer is input layer consisting of a group of processing units which are responsible for acceptance of data imported to the network. The input layer is connected with the hidden layer, which neither accepts input data directly nor produces output data directly. Different networks may consist of one or several hidden layers with the last layer connected with the input layer. Each input unit is connected with a hidden unit. All hidden units in the last of the hidden layers are connected with units in the input layer. There are one or several nodal points in the output layer which generate output data.

The connection weights between different nodal points of the neural network are different and constantly changing during the training. The course of network training is: to input a group of input data together with corresponding output data into the network which will adjust the weights between nodal points according to these data until the actual output is the same with or similar to the expected output. The network errors of each training group can be minimized with a substantial amount of input data. This method of weight changing between nodal points is very similar to the readjustment of the stimulus signals by synapsis of the brain: when the synapse is stimulated, the sum of the weights of the stimuli is transmitted to the karyoplast. When the sum of weights is bigger than the threshold value of this nerve cell, it will be activated and submit output. After stimulated by signals, a nerve cell only displays one of the two kinds of reactions: activated or not activated. The study process of the brain corresponds with the process of self-adjustment of the tightness between the neurons with the external stimulus information. The results of information processing of the brain are represented by the status of neuron.

Artificial neural network is widely engaged in radiology, but its application reports are rarely seen in China[10]. The neural network is a highly non-linear system which is endowed with parallel structure, parallel processing ability, separate knowledge storage ability, high fault toleration, self-adaptability and imagination. It is capable of learning. It is featured for its memory input mode, which is involved in the diagnosis for a new individual[11]. The BP network employed in this research is one of the most widely-used and popular common networks up to now[11]. Its main concept is to divide the course of learning into two stages: the first stage (positive direction diffusion process), in which input information is provided and processed through the input layer and hidden layer to get the actual output value of each unit; the second stage (negative direction process), in case the expected output value is not found in the output layer, then the difference (error) between the actual output and the expected output is calculated layer by layer recursively, the result of which shall be utilized for value adjustment.

To optimize and apply in use the trained network, the problems of its architecture and design have to be addressed. Input data shall be determined at the beginning stage of network design. Input nodal points can represent each datum so it is a must to find accurate data resources. If there are too many false or irrelevant data in the data resources, the proper training of the network[12, 13] will be obstructed. In this research, we have extracted 5 magnetic resonance image characteristics as input data to be input into the network. These 5 characteristics are meaningful in differentiating normal reactional enhancement and tumor residual. Some clinical characteristics and image characteristics that are not really relevant have been abandoned, including tumor pathological form, recheck time, edema situation, intactness of annular enhancement etc. Sources have it that the incidence rate of intracranial gliomas postoperative reactional enhancement is 65.7% (134/204), whereas that of tumor residual is 34.3%(70/204) [1]. The network is capable of learning and memorizing, and can memorize the data diffusion features. Therefore, if data with such rates are employed to train the network, it will overstudy reactional enhancement. Even though the network training is not successful and all are diagnosed as reactional enhancement, the correctness will be still as high as 65.7%. Therefore, the percents of reactional enhancement cases and tumor residual cases employed in this research are 50% respectively, which enables the network to study the magnetic resonance image characteristics of the two “justly”. Besides, the selection of nodal points of hidden layer neuron network based on BP is key to the performance of image network, so it is a must to select proper nodal points of the hidden layer. However, till present no effective approach has been found to determine the number of points of the hidden layer. This number is usually determined merely by experiences or experiments[12, 13]. Via numerous experiments, we have fixed the number as 6, and realized a comparatively sophisticated diagnosis ability of the network. According to this research, the diagnosis specificity of artificial neural network for a new individual is 85.7%, sensitivity 76.9%. To evaluate the diagnosis ability of the network, we have employed ROC method to compare it with the diagnosis ability of a senior radiologist. The ROC under curve area is Az=0.813±0.105, which is higher than that of the radiologist (Az=0.665±0.130). ROC has become one of the most widely used standard methods for the evaluation of 2 or more image diagnosis systems[14]. The proper training and application of neural network depends on the proper description of postoperative magnetic resonance image enhancement by an experienced doctor. Uniform standards shall be applied. The 5 indexes we have employed are easy to handle in practice, which ensures the consistency of image descriptions by network trainers and users. Thus, the best performance of the network is realized.

This research has proved that an artificial neural network is capable to differentiate the reactional enhancement of malignant intracranial gliomas in postoperative MR images and the enhancement caused by tumor residual. As a innovative differentiation approach, it requires further studies and researches to determine its applicability in practice, including the increase of training samples, image description standardization, effects to doctoral diagnosis, as well as parameter optimization. With improvements in these elements, the performance of the network will be optimized. As more doctors understand and accept the artificial neural network technology, it will for sure become a practical facilitate to assist radiologists to come out with a more accurate diagnosis.