Loss function for ineffective learning reduction in End-to-End deep image segmentation network

In image segmentation tasks based on deep learning methods, it is common that foreground pixels occur significantly more frequently than background pixels, and consequently bias the trained network towards them. In this paper, based on cost-sensitive learning, a design method of loss function for end-to-end image segmentation network is proposed, where two improvements are provided as follows: 1) Inspired by the conception of "hard examples mining", focal loss is introduced and extended to work for ineffective learning reduction. 2) Inspired by human visual systems, adaptive weights of receptive field are added to further consider the context information. In order to verify the validity and expansibility, the proposed method has been evaluated on severalmedicalimage datasets. The results show that the proposed method can improve the detection performance of the network for small objects, and obtain segmentation results that are more suitable for object contour.