Segmentation of Liver Cancer Pathology Images Based on Multi-scale Feature Fusion

Liver cancer is one of the common tumor diseases at present, and the segmentation of lesion area is the most important link in the early treatment of liver cancer. The high resolution of liver cancer pathology images, irregular borders of the lesion area, and uneven color distribution make traditional deep learning models not suitable for automatic segmentation of liver cancer pathology images. For this reason, we use the sliding window method to preprocess the dataset, which effectively avoids the loss of small parts of the input image caused by direct downsampling. In the network part, a multi-scale dilated convolutional neural network model is used to segment liver cancer pathology images. The multi-scale dilated convolutional neural network includes a dilated convolution feature pyramid structure with multiple receptive fields and an encode-decode structure with feature fusion. Through comparative experiments with the classic network models FCN, U-Net and SegNet on the liver cancer pathology images dataset, we discover that the model has a good effect on the segmentation of liver cancer pathology images, it achieved 85.2% and 86.3% in IOU and dice coefficient indicators. Compared with the classic network model, this model can fuse more different levels of image information.