Wnet ++: A Nested W-shaped Network with Multiscale Input and Adaptive Deep Supervision for Osteosarcoma Segmentation
Limei Shuai, Xin Gao, Jiajun Wang
Abstract
In this paper, a novel and more powerful architecture W-net++ was proposed based on two cascaded U-Nets and dense skip connections to realize the automatic and more accurate segmentation of osteosarcoma lesion in computed tomography (CT) images. In this network, multiscale inputs were applied to the architecture to recover the missing spatial details caused by multiple encoding and subsampling of the encoder; adaptive deep supervision was introduced to guide the multi-scale learning of the network to speed up convergence and improve the performance of the network; channel attention module (CAM) was incorporated to selectively emphasize the interdependent channel graphs and further improve the feature representation of the model. We have evaluated the proposed architecture and compared it with the-state-of-the-art methods by 5-fold cross validation on a home-built dataset of osteosarcoma CT images. Our experiments demonstrated that our method achieves an average DSC gain of 6.17 points, 1.91 points, 1.55 points over U-Net, U-Net++, MSRN, respectively.