基于深度学习的三维肿瘤及器官分割

Three-dimensional tumor and organ segmentation based on deep learning

针对三维医学图像中由于肿瘤或器官的形状、尺度差异较大导致分割精度较低的问题,提出一种端到端的三维全卷积分割模型。首先,设计空洞立方集成模块在不同分辨率阶段实现多尺度集成,增强复杂边界上的识别能力;其次,引入跨阶段上下文融合模块融合浅层和深层特征,促进收敛并更准确地定位目标对象;最后,解码器对来自编码器的特征进行拼接以实现分割。在脑肿瘤分割数据集上,平均Dice相似性系数值达到85.37%;在腹部器官分割数据集上,平均Dice相似性系数值达到83.99%。实验结果表明所提模型在三维肿瘤和器官的分割上具有较高精度。

In response to the challenge posed by the significant shape and scale variations of tumors and organs in threedimensional medical images,which often results in low segmentation accuracy,an end-to-end three-dimensional fully convolutional segmentation model is introduced.A dilated cubic integration module is designed to achieve multi-scale integration at different resolution stages,thereby enhancing the recognition capability on complex boundaries.Subsequently,a cross-stage context fusion module is incorporated to merge shallow and deep features,thereby facilitating convergence and more precise localization of the target objects.Finally,features from the encoder are concatenated by the decoder to realize segmentation.The average Dice similarity coefficients reach 85.37%on the brain tumor segmentation dataset and 83.99%on the abdominal organ segmentation dataset.Experimental results indicate that the proposed model exhibits high accuracy in three-dimensional tumor and organ segmentation.