Abstract: Accurate medical image segmentation is a necessary step in assisting disease diagnosis and
surgical planning. The automatic segmentation of liver tumors has always been a difficult problem
due to the blurred borders and the low contrast of abdominal organs. Aiming at the problem that the
traditional full convolutional network (FCN) achieves low accuracy in end-to-end segmentation, this
paper proposes a CT image liver tumor segmentation method based on convolutional multi-scale
fusion FCN. Firstly, the original CT image dataset is preprocessed by improving the contrast,
enhancement and denoising. Secondly, the designed FCN network is trained using the processed
dataset. Finally, a network model capable of accurately segmenting the liver tumor is obtained. The
experiment adopts a variety of evaluation indicators to evaluate the effectiveness of segmentation
results and makes comparison with a variety of common segmentation networks. The experimental
results show that the method proposed in this paper can accurately segment liver tumors of various
shapes and sizes in CT images, and the segmentation effect is good, which can provide a reliable
support for clinical diagnosis.

Key words: full convolutional network, liver tumor segmentation, deep learning, image segmentation;
convolution multi-scale fusion