The invention discloses a tumor prediction method for a PET-CT image based on a neural network. The tumor prediction method comprises the following steps: carrying out interpolation processing on a PET image and a CT image, wherein the lengths, widths and thicknesses of the two images are consistent; carrying out a coding process on the PET image and the CT image through a segmentation network, obtaining a PET feature image and a CT feature image, carrying out cascade combination, obtaining a final sample feature map, carrying out weighting on a PET feature part and a CT feature part of the final sample feature map, obtaining a segmentation model, and obtaining a tumor prediction image through the segmentation model; and performing a false positive removal process on the tumor prediction image through a 3D connected domain to obtain a tumor prediction result. The tumor is automatically segmented and identified by using the deep learning technology of the neural network, the suspected tumor tissue can be rapidly identified, the part is directly analyzed by a doctor, and finally, the diagnosis result is determined by the doctor, so that the working efficiency of a radiologist can be greatly reduced, and the tumor identification accuracy is improved.