Automated tumor characterization has a prominent role in the computer-aided diagnosis (CAD) system for the human brain. Despite being a well-studied topic, CAD of brain tumors poses severe challenges in some specific aspects. One such challenging problem is the category-based classification of brain tumors among glioma, meningioma, and pituitary tumors using magnetic resonance imaging (MRI) images. The emergence of deep learning and machine learning algorithms have addressed image classification tasks with promising results. But an associated limitation with the medical image classification is the small sizes of medical image databases. This limitation, in turn, limits the availability of medical images for training deep neural networks. To mitigate this challenge, we adopt a combination of convolutional neural network (CNN) features with support vector machine (SVM) for classification of the medical images. The fully automated system is evaluated using Figshare open dataset containing MRI images for the three types of brain tumors. CNN is designed to extract features from brain MRI images. For enhanced performance, a multiclass SVM is used with CNN features. Testing and evaluation of the integrated system followed a fivefold cross-validation procedure. The proposed model attained an overall classification accuracy of 95.82%, better than the state-of-the-art method. Extensive experiments are performed on other MRI datasets for the brain to ascertain the improved performance of the proposed system. When the amount of available training data is small, the SVM classifier is observed to perform better than the softmax classifier for the CNN features. Compared to transfer learning-based classification, the adopted strategy of CNN-SVM has lesser computations and memory requirements.

自动化的肿瘤表征在人脑的计算机辅助诊断（CAD）系统中具有重要作用。尽管这是一个经过充分研究的话题，但脑肿瘤的CAD在某些特定方面提出了严峻的挑战。这样一种具有挑战性的问题是使用磁共振成像（MRI）图像对脑胶质瘤，脑膜瘤和垂体瘤中的脑肿瘤进行基于类别的分类。深度学习和机器学习算法的出现解决了图像分类任务，并取得了令人鼓舞的结果。但是医学图像分类的相关限制是医学图像数据库的小尺寸。这种限制反过来又限制了用于训练深度神经网络的医学图像的可用性。为了缓解这一挑战，我们采用卷积神经网络（CNN）功能与支持向量机（SVM）的组合来对医学图像进行分类。全自动系统使用无花果开放的数据集，其中包含针对三种类型的脑肿瘤的MRI图像。CNN旨在从大脑MRI图像中提取特征。为了增强性能，将多类SVM与CNN功能一起使用。集成系统的测试和评估遵循五重交叉验证程序。所提出的模型获得了95.82％的总体分类精度，优于最新方法。在针对大脑的其他MRI数据集上进行了广泛的实验，以确定所提出系统的改进性能。当可用训练数据量较小时，对于CNN功能，SVM分类器的性能要优于softmax分类器。与基于迁移学习的分类相比，采用的CNN-SVM策略具有较少的计算和内存需求。