Overall, with the novelty of DL, there were merely 273 SCI papers reviewed in this study of DL and spine, with a major increase in these publications since 2020. In our study, China was the country with the maximal quantity of publications (88, 32.2%), while USA was the country with the highest H-index of 15. According to top 10 institutions with publications, Chinese and American institutions came forward once again, and Sun Yat Sen University was the institution with the most publications (14, 5.1%). Recently, publications about applications of DL in spine significant increased. However, different from AI, the development of those techniques in spine are still in infancy on account of the fact that the notion of DL was first proposed by Hinton[25] in 2006 and its wide application was much later. As a subset of ML, the sample size determines the quality of the output results[26], and deep learning should be fed with raw data and develops its own representations needed for pattern recognition. China and USA are two largest countries over the world, and their clinical database may be the basis of their productive publications, which was far more than other countries/regions. However, though quantity plays a critical role in the application of ML in the field of spine, quality also merits our attention. As shown in Fig. 3 and Fig. 5, the cooperation between different countries or institutions remains deficient. Therefore, we appeal for a standard dataset and data sharing over the world, and only by this way, will DL function more efficient and proficient.
The top ten journals and their relations were shown in Table 3 and Fig. 7, whose coverage could be classified as four aspects: spine topic (European spine journal, Spine journal and Global spine journal), algorithm and computational topic (Ieee access, Computer methods and programs in biomedicine), radiographic topic (Medical image analysis, European radiology, and Journal of digital imaging) and comprehensive topic (Scientific reports and Journal of clinical medicine). As the interdisciplinary combination between DL and spine deepens, it presented an inevitable problem for the authors to choose reasonable journals. And we hope our study could give some recommendation for authors of different disciplines. Research category analysis was shown in Table 4, whose distribution was similar to the top ten journals. Radiology Nuclear Medicine Medical Imaging won the first place with 77 publications (H-index 16), while Clinical Neurology and Orthopedics ranked the second and third place. The application of DL mainly initials in imaging[27]. As the clinical practice of spine primarily depends on radiography, DL acts as an initial drive in the screening and diagnosis in the spinal disorders, such as scoliosis[28], disc herniation[29] and spinal tumor[30]. And more novel and comprehensive deep learning algorithm may realize the goal of automation detection and diagnosis one day. However, the application of DL on the treatment or guidance of spine still remains scarce. Saravi et al[31] ever performed deep learning algorithms to predict prolonged length of stay after lumbar decompression surgery, and Yagi et al[32] predicated major complications 2 years after corrective spine surgery for adult spinal deformity using DL model. All the novel technologies are designed to solve the clinical problem, and therefore we believed the future studies will aim for the application of ML in more complex problems in the field of spinal diseases, rather than merely in radiographical aspect.
VOSviewer was adopted to analyze 113 keywords with at least 10 occurrence times, and Fig. 8 showed the consequence presented three clusters which consisted of “segmentation”, “area”, and “neural network”. Spine, as the axial skeleton, can be divided into cervical, thoracic, lumbar and sacral parts, and meanwhile, each part constitutes of vertebrae, disc and related level of spine cord[33]. Precise segmentation and anatomical identification of the spine provides the basis for automatic analysis of the spine[34]. To further localize the lesion of spinal diseases, segmentation is a necessary step in the application of ML in spine diseases. In the top ten cited articles, Lessmann et al[34], Li et al[35], Horng[36], and Al Arif[37] emphasized the importance of segmentation. And reconstruction becomes a way to realize the output of spine, while its accuracy remains improvement[38]. DL methods suffer from the “black box” problem: input is supplied to the algorithm and an output emerges, but it is not exactly clear what features are identified or how they inform the model output[39]. Therefore, the examination and test of accuracy plays a critical role. Area under receiver operating characteristic curve, agreement examination and numerous statistical methods were adopted to assure the accuracy of the DL models. Neural network won the most frequent occurrence among the 113 key word, which is concrete algorithm of deep learning. Moreover, convolutional neural network (CNN) has been widely applied in the image problems[40–42]. Our study indicated CNN was one of the keywords with the greatest outburst intensity, which was consistent to most studies related to our topic that focus on the field of radiographic image. “magnetic resonance image” and “lumbar” were the keywords with the longest use (2017–2020), which indicated the fact that most studies performed their research on MRI and lumbar region[13, 43–46]. And “agreement” and “automated detection” were the most popular keywords from 2020 to 2022, which implied that automated detection would become a trend of spine in the future[47–49].
Future Prospect
Our study showed the trend of application of deep learning in spine in this decade, the countries, institutions and journals with most number of publications and citations were listed, and the co-occurrence of the key words were analyzed. However, to favorably apply this novel technology in the spinal field, there still needs persistent efforts and innovation. Firstly, in view of the attribute of DL, standard database and shared huge datasets should be constructed, which depends on closer cooperation between different institutions, regions and countries. Besides, the privacy protection is also essential. Secondly, spine is not only a structure of our body, its related diseases and clinical practice are more remarkable than itself for clinicians and patients. Therefore, regardless of screening and diagnosis, the application of DL should also be widely promoted in surgical decision making, intraoperative manipulation, prognosis prediction and rehabilitation of spinal disorders in the near future. What’s more, as the black-box principle of DL, the interpretability of networks should be improved. Accuracy examination and algorithm optimization will be worthwhile before clinical practice.