A Novel Method Based on Convolutional Features with Non-Iterative Learning for Brain Tumor Classification

Brain tumor is a cluster of abnormal and uncontrolled cells growth, leading to a short life expectancy in their highest grade. Accurately and timely clinical diagnosis of such tumors play critical role in treatment planning and patient care. Various image techniques such as Computed Tomography (CT), Ultra-Sound image, Magnetic Resonance Imaging (MRI) and biopsy are used to evaluate brain tumors. Among stated four, MRI is most common used non-invasive technique, however the key challenge with these images is the low-level visual information captured by MRI machines, that needs highlevel interpretation by experienced radiologist. The manual interpretation is a tedious, challenging, and erroneous task. In this paper, we propose a novel Convolutional Feature based Euclidean Distance (ConFED) method for faster and more accurate tumor classification. The method consists of convolutional features and Euclidean distance based one step learning. The proposed method is evaluated on Contrast-Enhanced Magnetic Resonance Images (CE-MRI) benchmark dataset. Proposed method is more generic as it does not use any handcrafted features, requires minimal preprocessing, and can achieve average accuracy of 97.02% using five-fold cross-validation. Extensive experiments, along with statistical tests, revealed that the proposed method has outperformed state-of-the-art classification methods on the CE-MRI dataset

脑肿瘤是一类异常且不受控增殖的细胞团，在最高分级时会大幅缩短患者的预期寿命。对这类肿瘤开展准确且及时的临床诊断，对于治疗方案规划与患者照护至关重要。当前临床常用计算机断层扫描（Computed Tomography, CT）、超声成像、磁共振成像（Magnetic Resonance Imaging, MRI）以及活检等多种影像技术来评估脑肿瘤。在上述四种手段中，MRI是最常用的非侵入式检查技术，但这类图像面临的核心挑战在于磁共振设备采集的视觉信息层级较低，需由经验丰富的放射科医师进行高阶解读。人工解读不仅工作繁琐、难度极大，且易产生误差。本文提出一种基于卷积特征的欧氏距离（Convolutional Feature based Euclidean Distance, ConFED）方法，以实现更快且更精准的肿瘤分类。该方法融合卷积特征与基于欧氏距离的单步学习机制。所提方法在对比增强磁共振成像（Contrast-Enhanced Magnetic Resonance Images, CE-MRI）基准数据集上进行了验证。该方法具备更强的通用性：无需使用任何手工设计特征，仅需极少量的预处理工作，且通过五折交叉验证可达到97.02%的平均分类准确率。大量实验结合统计检验结果表明，在CE-MRI基准数据集上，所提方法的性能优于当前领域内最优的分类算法。