AI-based cervical lesion segmentation in colposcopy images has significant potential in improving screening efficiency and accuracy. However, most current cervical lesion segmentation algorithms are insufficient for rapid image segmentation in mass screening due to heavy parameters and complex framework. Therefore, a lightweight algorithm segmentation for cervical lesion real-time screening system is urgently needed. In this paper, a novel lightweight LSIL + region segmentation framework termed Light-MDDNet is proposed and deployed, which uses the encoder-decoder architecture. In encoder stage, the first layer of MobileNetV2 module outputs shallow features which tend to be lost during multi-layer feature extraction. We also utilize the Denseaspp module to extract deeper semantic information. In the decoder stage, a multi-scale feature fusion (MFF) module is used to fuse multi-scale features. Finally, the algorithm is deployed and tested on the JETSON ORIN NX edge device in cervical lesion segmentation screening system. The experiments on 971 LSIL + colposcopy images of lesions processed with acetic acid show that our proposed model outperforms some state-of-the-art segmentation networks, with a pixel mean pixel accuracy (MPA) of 94.96% and an average speed per image of 19.60ms. After deployment on the mobile terminal, the segmentation accuracy of the model almost unchanged and the interference speed reduces to 31.57ms per image. The Light-MDDNet network achieves the best balance of accuracy and speed in cervical lesion segmentation, showing great potential for the deployment in the mass screening of cervical lesion.