龋齿是一种常见的口腔疾病，早期龋仅仅为牙齿硬组织的脱钙，往往没有明显的缺损和症状，若及早发现并及时采取预防措施，脱钙的牙齿还可以再矿化。如果龋坏进一步加重，治疗过程将十分复杂。由于早期龋齿一般症状为轻微脱矿，在临床X光牙片上多表现为不明显的灰度变化，医生肉眼阅片有一定难度。而计算机处理和区分灰度图像的优势使其逐渐成为医生肉眼诊断的有力辅助工具。所以，建立一套计算机龋齿诊断系统来辅助医师的诊断有十分重要的现实意义。目前，针对早期龋齿诊断的研究较少，主要有两种方法：一种通过制定规则来判断有无龋齿，另一种通过提取特征代入模式识别分类器中来判断，但其效果都不尽如人意。因此，本文针对临床浅龋的特点，研究和搭建了一套龋齿计算机辅助诊断平台。该辅助诊断平台主要分两部分：首先利用了分水岭算法对X光片的直方图进行分水岭变换。变换后的在各个分水岭之间寻找阈值，并根据阈值对原图像重新进行赋值。另外，本文还借用了遥感图像中常用的伪彩色变换思想，另外生成一幅伪彩色图像，通过图像增强将龋坏部位和正常部位区分开来。 之后，对怀疑区域提取特征，并将其输入支持向量分类机模型中来进一步确定有无龋齿。通过实验，支持向量机的分类准确率高于常用的模式识别分类机——神经网络分类器约10个百分点。为了提高识别率，本文对支持向量机的核函数进行了改进，修改后的模型识别准确率高于原始模型近9个百分点，通过交叉校验选择模型参数使得模型最终分类准确率再提高3个百分点，为77.1428%，高于高年资医生肉眼阅片准确率4个百分点（73%），高于普通医生15个百分点（62%）。可以为医生的临床诊断提供较为准确的辅助意见。

Some of the initial caries appear as an unnoticeable loss of mineral substance and the X-ray image corresponding to it only has a slight change in gray-level values, it is difficult for dentists, especially for those inexperienced doctors, to make exact diagnosis. For computers have superior visual distinction than human eyes, when it comes to medical treatment, a computer-aided diagnosis system, as an auxiliary tool, could give doctors constructive suggestions in order to prevent misdiagnosis or missed diagnosis. Computer-aided technology has been used for clinical diagnostic applications. However, researches or products specializing in computer-based caries detection are much rarer. Considering the merits of both image process and pattern recognition technologies, we combined these two methods together to make full use of their own advantages and propose a computer-aided caries detection system. First, we applied watershed algorithm on the histogram of the dental X-ray images. Using the thresholds acquired from the transformation to reallocate grey-level values of the original image. This enhancement process could amplify the decayed area; therefore, it is easier for dentists to discriminate the caries from normal teeth. In addition, we also generate a pseudo-colored image to further enlarge the difference between lesion and normal area.Then, we select Support Vector Machine as the prediction model. In order to increase the diagnostic accuracy, we modified the kernel function of SVM. Experiments show that the testing accuracy has improved approximately 9 percent and the diagnosis accuracy is 4 percent higher than does the professional dentist and 15 percent higher than does the mean of three ordinary dentists. Therefore we could safely draw the conclusion that our system could give dentists an effective suggestion to avoid checking failures.