The invention relates to an abdominal muscle labeling method and device based on deep learning. The method comprises the following steps of collecting the abdominal CT image data containing a third lumbar vertebra; marking a third lumbar vertebra position and a muscle group position, wherein four muscle group areas are marked as 1, 2, 3 and 4 respectively, and other areas are marked as 0; generating a label image corresponding to the original CT image, wherein the value of each pixel in the label image is one of {0, 1, 2, 3 and 4}; utilizing the labeled CT image to train a segmentation model,dividing pixels in the CT image into five classes by the segmentation model, and enabling the five classes of pixels to respectively correspond to the labels 0, 1, 2, 3 and 4 in the second step; segmenting the muscle group to obtain the label prediction corresponding to each pixel position in the image; and based on the muscle group segmentation result, calculating the muscle area and the image omics characteristics of the muscle. The device includes the related modules that implement the method. By utilizing the method, the parameters related to the nutrition assessment can be simply, conveniently, quickly and accurately extracted.