Spine segmentation for magnetic resonance (MR) images is important for various spinal diseases diagnosis and treatment, yet is still a challenge due to the inter-class similarity, i.e., shape and appearance similarities appear in neighboring spinal structures. To reduce inter-class similarity, existing approaches focus on enhancing the semantic information of spinal structures in the supervised segmentation network, whose generalization is limited by the size of pixel-level annotated dataset. In this paper, we propose a novel detection-guided mixed-supervised segmentation network (DGMSNet) to achieve automated spine segmentation. DGMSNet consists of a segmentation path for generating the spine segmentation prediction and a detection path (i.e., regression network) for producing heatmaps prediction of keypoints. A detection-guided learner in the detection path is introduced to generate a dynamic parameter, which is employed to produce a semantic feature map for segmentation path by adaptive convolution. A mixed-supervised loss comprised of a weighted combination of segmentation loss and detection loss is utilized to train DGMSNet with a pixel-level annotated dataset and a keypoints-detection annotated dataset. During training, a series of models are trained with various loss weights. In inference, a detection-guided label fusion approach is proposed to integrate the segmentation predictions generated by those trained models according to the consistency of predictions from the segmentation path and detection path. Experiments on T2-weighted MR images show that DGMSNet achieves the state-of-the-art performance with mean Dice similarity coefficients of 94.39% and 87.21% for segmentations of 5 vertebral bodies and 5 intervertebral discs on the in-house and public datasets respectively.
Keywords: Deep learning; Ensemble learning; Mixed-supervised segmentation; Spine.
Copyright © 2021 Elsevier B.V. All rights reserved.