Abstract: This paper mainly studied the intelligent detection of lesions about LCI (Linked Color Imaging) images in the medical field. The difference of shapes and colors between lesion regions and non-lesion regions in LCI images with symptoms of inflammations and early cancers could be used as a classification principle. However, due to the low degree of discrimination between region boundaries, detected lesion regions were usually not consistent with actual lesion regions. In order to obtain accurate and detailed lesions, this paper based on the fully convolutional network, then used SVM(Support Vector Machine) function as the loss function and trained models for the limitation of blurred boundaries to achieve pixel-level classification of images. Based on the classification result whether the pixel belongs to lesion regions or not, the boundaries between lesion regions and non-lesion regions could be determined. The proposed algorithm was compared with FCN(Fully Convolutional Network) and the traditional semantic segmentation algorithm GrabCut, The experimental results showed that the proposed algorithm outperformed other algorithms with accuracy 94%, and finished the detection on a single image in 0.5 s averagely, which could help doctors to quickly diagnose the disease and have great clinical significance.