The invention discloses a breast tumor segmenting method based on a multistage converting network, and relates to the technical field of medical image treatment. The breast tumor segmenting method comprises the following steps of firstly, enabling experts to classify tumor ultrasound images according to the size of tumors, performing repeated training according to classification results, and constructing a volume assessment network; then studying the characteristics of large tumors, performing amplifying treatment on small tumor ultrasound images to generate high-quality large tumor images, and constructing an image converting network; and finally, performing segmenting training on the large tumor ultrasound images and amplified small tumor ultrasound images, and after segmenting is completed, zooming the small tumor ultrasound images according to the same ratio, so as to complete the construction of whole-convolution neural network. According to the breast tumor segmenting method disclosed by the invention, segmenting can be performed according to the size of tumors in the tumor ultrasound images, and the segmenting precision and the segmenting efficiency of the small tumors can be especially improved. The invention also provides a breast tumor segmenting device based on a multistage converting network. The breast tumor segmenting device is combined with the breast tumor segmenting method, so that the segmenting of the small tumor ultrasound images is better completed.