3D multi-scale feature extraction and recalibration network for spinal structure and lesion segmentation

Background Automatic segmentation has emerged as a promising technique for the diagnosis of spinal conditions. Purpose To design and evaluate a deep convolution network for segmenting the intervertebral disc, spinal canal, facet joint, and herniated disk on magnetic resonance imaging (MRI) scans. Material and Methods MRI scans of 70 patients with disc herniation were gathered and manually annotated by radiologists. A novel deep neural network was developed, comprising 3D squeeze-and-excitation blocks and multi-scale feature extraction blocks for automated segmentation of spinal structure and lesion. To address the issue of class imbalance, a weighted cross-entropy loss was introduced for training. In addition, semi-supervision segmentation was accomplished to reduce annotation labor cost. Results The proposed model achieved 77.67% mean intersection over union, with 9.56% and 11.11% gains over typical V-Net and U-Net respectively, outperforming the other models in ablation experiments. In addition, the semi-supervision segmentation method was proven to work. Conclusion The 3D multi-scale feature extraction and recalibration network achieved an excellent segmentation performance of intervertebral disc, spinal canal, facet joint, and herniated disk, outperforming typical encoder-decoder networks.