【摘要】 针对现有的全卷积网络处理脑肿瘤分割任务时网络参数量大、计算困难的问题,提出了一种结合随机森林(Random Forests, RF)和密集连接网络(DenseNet)的方法.方法分为粗分割和精细分割两部分.粗分割在下采样的脑磁共振图像(MRI)上用增强RF初步分割出肿瘤.精细分割依据粗分割得到原始MRI的感兴趣区域,用改进的DenseNet精细分割感兴趣区域为坏死、水肿、非增强肿瘤、增强肿瘤和其它.最后还原精细分割,形成原始MRI的分割结果.实验结果表明,提出的方法完成一次分割所需时间不到6 s,保证分割精度的同时降低了参数数量和计算量.更多还原

【Abstract】 The existing full convolutional networks have problems that have too many parameters and huge computational consumption in training when processing brain tumor segmentation tasks. To address the problems, a method combining random forests and DenseNet is proposed. The method is divided into rough segmentation and fine segmentation. The rough segmentation provides the preliminary prediction results, which is done by enhanced RF method on the downsampled MRI. The regions of interest(ROI) are extracted from original MRI based on the preliminary prediction results. The fine segmentation processes the ROI and outputs the prediction results including necrosis, edema, non-enhancing tumor, enhancing tumor and anything else by improved DenseNet. Finally, the prediction results are restored to original MRI segmentation results. Experiments show the proposed method takes less than 6 seconds to complete a segmentation. This indicates that the proposed method ensures segmentation accuracy while reducing the number of parameters and calculation requirements.更多还原