基于目标检测的医学影像分割算法

Medical Image Segmentation Based on Object Detection

提出了一种基于目标检测的医学影像分割算法。目标检测网络可以从医学影像中获取到感兴趣目标(一般为器官或病变组织)的精确位置信息,根据位置信息对医学影像进行裁剪得到感兴趣目标的图像块。以感兴趣目标图像块的关键区域作为参考对象,以该图像块作为搜索对象,孪生网络可以获取到图像块中的每一个像素点被分类为关键像素点的概率,根据预先设定好的概率阈值可得到感兴趣目标的分割结果。使用肝脏(LITS17)和肾脏(KITS19)两个数据集对上述医学影像分割算法进行评估。实验结果表明该分割方法可以较准确地对医学影像中的感兴趣目标进行分割。

This paper proposes a medical image segmentation algorithm based on object detection.With the help of an object detection network,the precise locations of the object of interest(usually an organ or diseased tissue)in medical images are obtained.And the medical images are cropped according to locations obtained by the object detection network to gain the image patches of objects of interest.By taking the key center area of the image patch of objects of interest as the examplar image and this image patch as the search image,the pixel-wise classification of this image patch is achieved by one Siamese network with the help of a pre-defined threshold.Two datasets(LITS17 and KITS19)are used to evaluate our algorithm.And experiments show that the proposed method can accurately segment organs or diseased tissues in medical images.