Accurate segmentation of red blood cell (RBC) from blood smear images is an important technique and a difficult problem, mainly because RBCs often overlap and have no distinct boundaries. To solve this problem, a deep learning network called NODE-UNet++ is proposed, which is based on U-Net++ and neural ordinary differential equations (NODE). It is mainly used for pre-segmentation of RBCs, and then the marker watershed algorithm is adopted to segment clustered RBCs from blood smear images. Firstly, an image is clipped and labeled to highlight the region to be segmented. Then, a new semantic segmentation architecture NODE-UNet ++ is applied for pre-segmentation of the preprocessed image to obtain the probability grayscale image. Finally, the marker watershed method is used to separate the clustered RBCs in the grayscale image to obtain final RBC segmentation result. The experimental results show that the Dice similarly coefficient is 96.89%, the mean pixel accuracy is 98.97%, and the mean intersection over union is 96.33%. Segmentation results of different blood smear images show that the proposed method can extract each RBC efficiently and accurately to meet the requirements of subsequent RBC image processing.