Table 1.

Publications on spine image analysis based on “SpineWeb” (years 2019–2021).

Study/ Number of Data	Algorithms Applied	Objectives	Outcome Presentation
Liansheng W. [93] 707 spinal AP X-ray images	U-net	Top eight methods from twelve teams (including intuition, workflow, and implementation). Experimental results show that, overall, the best performing method achieved an asymmetric mean absolute percentage (SMAPE) of 21.7%.	Quantitative measurement of the spine.
Liyan L. [116] 895 axial spine MRI images from 143 patients	OSBP-Net, IPDC, and IICR	Applied to the output of the SFEs, taking into account that the activated regions in the feature maps of two paths should be theoretically different.	The prediction results, comparison with many other CADq models
Shen Z. [123] 450 MRI scans	Can-See is a two-step detection framework: (1)A hierarchical proposal network (HPN) to perceive the existence of the vertebrae. (2)A category-consistent self-calibration recognition (CSRN) network used to classify each vertebra and to refine their bounding boxes.	Category-consistent self-calibration recognition system (Can-See) used to accurately classify the labels and precisely predict the bounding boxes of all vertebrae with improved discriminative capabilities for vertebrae categories and self-awareness of false-positive detections.	Can-See achieves high performance (testing accuracy reaches 0.955) and outperforms other state-of-the-art methods.
Zhongyi H. [113] 253 clinical patients	Neural-symbolic learning (NSL) framework	Compares the semantic segmentation ability of a neural symbolic learning framework (NSL) with several state-of-the-art semantic segmentation networks (FCN, SegNet, DeepLabV3+, U-Net, Spine-GAN, GN-SGR, AGN-SGR, and AGN-DN).	NSL can directly generate radiologist-level diagnosis reports (using two steps) in spine radiology.
Dong Z. [118] 240 subjects	Sequential conditional reinforcement learning (SCRL). SCRL coordinates three major components (AMRL, Y-Net and FC-ResNet)	Propose a sequential conditional reinforcement learning network (SCRL) to tackle the simultaneous detection and segmentation of VBs from MR spine images.	SCRL achieves accurate detection and segmentation results, where on average, the detection IoU is 92.3%, segmentation dice is 92.6%, and classification mean accuracy is 96.4%.
Yanfei H. [117] 200 subjects	MMCL-Net: (1)The densely dilated ResNet, (2)The deep convolution level set module, (3)The instance feature merge module combines the global features extracted by DDRN and the local features obtained by segmentation	Novel end-to-end multi-task multi-structure correlation learning network (MMCL-Net) for the detection, segmentation and classification (normal, slight, marked and severe) of three types of spine structure: disc, vertebra and neural foramen simultaneously	MMCL-Net achieves high performance with a mAP of 0.9187, a classification accuracy of 90.67%, and a dice coefficient of 90.60%.
Liyan L. [116] 895 axial spine MRI images from 143 subjects	Dense enhancing network (DE-Net)	Dense enhancing network (DE-Net), which uses the dense enhancing blocks (DEBs) as its main body.	All deep learning models obtain very small prediction errors, and the proposed DE-Net with CSDPR acquires the smallest error among all methods.
Ranran Z. [115] 407 subjects	Multi-task relational learning network (MRLN)	A dilation convolution group is used to expand the receptive field, and LSTM (long short-term memory) to learn the prior knowledge of the order relationship between the vertebral bodies.	The accurate segmentation, localization and identification of vertebrae.
Jiawei H. [112] 320 axial lumbar MRIs	BS-ESNet	For the first time: (1)segmentation of the multiple paraspinal muscles and other major spinal components on axial lumbar MRIs simultaneously at both upper and lower spinal levels is achieved. (2)Boundary sensitive network provides a novel segment-then-detect workflow, which is robust to unclear organ boundaries and can further simplify multi-organ detection as an end-to-end trainable process; (3)Explicit saliency-aware network provides an elaborately designed architecture, which can utilize detection b-boxes to automatically correct and enhance segmentation features in an explicitly supervised manner and facilitates the adaptation of variable precise anatomical structures.	Proposal an explicit saliency-aware learning framework for segmentation of paraspinal muscles at varied spine levels.
Heyou Ch. [114] 292 MRI scans	A spatial graph convolutional network (GCN)	The proposed method is trained in an end-to-end.	Method achieves high performance (89.28 ± 5.21) of IDR and (85.37 ± 4.09%) of mIoU) from arbitrary input images.
Shen Z. [124] none	Adversarial recognition (FAR) network	Network to accurately perform spondylolisthesis grading by excellently detecting critical vertebrae without the need for locating landmarks.	Training accuracy: 0.9883 ± 0.0094, testing accuracy: 0.8933 ± 0.0276 for MRI images of different modalities, which can be attributed to the excellent critical vertebrae detection (detection mAP75 for training: 1 ± 0, for testing: 0.9636 ± 0.0180, and IoU (intersection-over-union) ≥ 0.9/0.8 for most detections with their corresponding ground truth in the training/testing dataset).
Liansheng W. [88] 526 X-rays	MVE-Net	Proposed multi-view extrapolation net (MVE-Net) that provides accurate automated scoliosis estimation in multi-view (both AP and LAT) X-rays.	Experimental results on 526 X-rays show 7.81 and 6.26 circular mean absolute error in AP and LAT angle estimation, which shows the MVE-Net provides an accurate Cobb angle estimation in multi-view X-rays
Shen Z. [123] none	Faster adversarial recognition (FAR)	Proposed faster adversarial recognition (FAR) network to accurately perform spondylolisthesis grading by excellently detecting critical vertebrae without the need for locating landmarks.	training accuracy: 0.9883 ± 0.0094, testing accuracy: 0.8933 ± 0.0276 for MRI images of different modalities, which can be attributed to the excellent critical vertebrae detection (detection mAP75 for training: 1 ± 0, for testing: 0.9636 ± 0.0180, and IoU (intersection-over-union) ≥ 0.9/0.8 for most detections with their corresponding ground truth in the training/testing dataset).
Shumao P. [111] MR images of 215 subjects	Cascade amplifier regression network (CARN)	Proposed novel cascade amplifier regression network (CARN) with manifold regularization including local structure-preserved manifold regularization (LSPMR) and adaptive local shape-constrained manifold regularization (ALSCMR) to achieve accurate direct automated multiple indices estimation.	Proposed approach achieves impressive performance with mean absolute errors of 1.22±1.04 mm and 1.24 ± 1.07 mm for the estimation of 30 lumbar spinal indices of the T1-weighted and T2-weighted spinal MR images, respectively.