Comparative Study of AI Models for Pathological Fractures Identification in Spinal Metastases

Comparative Study of AI Models for Pathological Fractures Identification in Spinal Metastases | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Comparative Study of AI Models for Pathological Fractures Identification in Spinal Metastases Yaowen Zhang, Zihan Qin, Nan Bao, Hong Li, Zhen Liu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5364269/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Bone metastasis is a common complication in most malignant tumors, with the spine being a frequent site of metastasis. Cancerous spinal bone metastases can easily lead to pathological fractures, resulting in pain, neurological deficits, and a deterioration in the quality of life for patients. Therefore, early identification of pathological fractures in spinal metastases is crucial for improving the survival quality of patients with advanced cancer. This retrospective study collected CT imaging data from 140 patients with spinal bone metastases and resultant fractures, including 77 male and 63 female patients, with an age range of 34 to 79 years. The study began by preprocessing the CT images and had experienced radiologists annotate the fractures. Subsequently, deep learning-based models, single-stage YOLO-v5, YOLO-v7, and YOLO-v8, as well as the two-stage Faster R-CNN, were constructed and analyzed to detect pathological fractures caused by cancerous spinal metastases. Experimental results indicate that these AI models achieved favorable detection performance, with YOLO-v8 demonstrating the best performance on an independent testing set, achieving a precision of 0.999, a recall rate of 0.998, and a mean average precision (mAP) of 0.991. The vertebral fracture identification model based on YOLO-v8 can provide strong technical support for clinical diagnosis. Health sciences/Health care/Diagnosis Health sciences/Health care/Medical imaging/Bone imaging vertebral fracture fracture identification spinal cancer metastases deep learning YOLO Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Introduction According to data statistics from the World Health Organization (WHO), approximately 20 million new cancer cases were reported globally in 2022, with nearly 10 million cancer-related deaths [ 1 ]. Bone metastasis is a common complication in most malignant tumors, with the spine being a frequent site of metastasis [ 2 , 3 ]. Statistics show that about 70% of cancer patients will develop spinal metastases [ 4 ], and 10%-20% of these cases will experience concurrent symptoms, leading to pain, neurological deficits, and a deterioration in quality of life [ 5 ]. Moreover, 10% of patients may suffer from spinal cord compression, which can result in potentially permanent disability [ 6 ]. The bone destruction caused by spinal metastases mainly manifests in three types: osteoblastic, lytic, and mixed [ 7 ]. Data indicate that lytic spinal metastases occur more frequently, accounting for about 70% of cases [ 8 ]. The incidence of pathological fractures due to bone metastases ranges from 10–29% [ 2 ]. As tumors grow and affect the vertebral body or compress the spinal cord and nerve roots, symptoms such as pain, bowel dysfunction, sensory and motor dysfunction of the lower limbs, and even paralysis may arise [ 9 ]. Therefore, early detection and prediction of pathological fractures in spinal metastases are crucial for improving the quality of life of patients with advanced cancer [ 10 – 12 ]. In recent years, radiomics technology has rapidly developed, which uses computer analysis to extract features from medical images, transforming them into high-dimensional data for analysis [ 13 , 14 ]. This technology has the potential to improve disease diagnosis and prognosis. Deep learning can autonomously learn data features as part of machine learning, reducing manual processing. Integrating deep learning with machine learning and radiomics has shown significance in cancer research [ 15 , 16 ]. Muehlematter et al. [ 17 ] employed machine learning and texture analysis techniques to evaluate consecutive CT scan images of 58 patients with spinal insufficiency fractures in spinal fracture detection. The study found that a Support Vector Machine (SVM) classifier combined with 29 texture parameter features achieved a high AUC (Area Under Curve) of 0.97 in identifying patients with spinal insufficiency. Oh et al. [ 18 ] studied 84 lung cancer patients with proximal femoral bone metastases who underwent CT scans and were followed up for at least three months. They extracted and standardized 16 clinical and radiological features and then trained these features using five supervised learning classifiers. The study revealed that the Gradient Boosting (GB) algorithm had the best predictive performance, with an AUC of 0.80 ± 0.14. Additionally, Wang et al. [ 19 ] collected clinical and radiological data from 1143 patients with bone metastases and used machine learning techniques to predict Skeletal-Related Events (SREs). They found that an SVM classifier with eight variables achieved an accuracy of 97.1% in predicting SREs. Ramachandran et al. [ 20 ] utilized a single-stage YOLO (You Only Look Once) network structure based on the DetectNet architecture for detecting pulmonary nodules in lung CT images. This method achieved a meager false-positive rate, demonstrating high sensitivity and accuracy. Xu et al. [ 21 ] designed an innovative cascaded convolutional neural network, W-Net, to learn and extract critical features from multi-modal medical images autonomously. These research outcomes suggest that deep learning-based object detection technologies have potential applications in medical diagnostics. Therefore, this study aims to explore the use of deep learning networks for the target detection of pathological fractures caused by cancerous spinal bone metastases. Results In this study, models such as Faster R-CNN, YOLO-v5, YOLO-v7, and YOLO-v8 were constructed on the training dataset. Figure 1 presents the training-loss curves for the Faster R-CNN, YOLO-v5, YOLO-v7, and YOLO-v8 models. It can be observed that all four curves decrease over iterations and stabilize, demonstrating that the models were trained effectively. Figure 2 shows the prediction results of YOLO-v8, where the top row is the label images and the bottom row is the prediction results. Table 1 shows the precision, recall, and mAP corresponding to the prediction results of the four models on the test set. Table 1 Table 1 . Prediction performance on the testing set. Model Precision Recall mAP FasterR-CNN 0.952 0.964 0.960 YOLO-v5 0.983 0.990 0.989 YOLO-v7 0.979 0.977 0.980 YOLO-v8 0.999 0.998 0.991 Discussion This study utilized four object detection models to detect fractures on CT images of cancer patients with spinal bone metastases. The experimental results indicate that all four models achieved satisfactory detection after thorough learning, training, and appropriate parameter adjustments. If mAP is used as the critical evaluation index, we can observe that YOLO-v8 has a clear advantage over other networks. Figure 3 shows the comparison results between YOLO-v8 and YOLO-v7, YOLO-v5, and Faster R-CNN. The mAP comparison between YOLOv8 and YOLOv7 shows that the former has a significantly better convergence speed than the latter. From the initial few epochs, the performance of the YOLOv8 model starts to outperform YOLOv7 significantly, and the mAP value of YOLOv8 reaches 0.991, slightly higher than that of the comparative network. In the comparison between YOLOv8 and YOLOv5, it can be seen from the graph that the mAP curves of the two models generally remain consistent. Initially, the two curves change more or less in unison, but at around 30 epochs, the performance of the YOLOv8 model begins to surpass YOLOv5. In the end, the mAP value of the YOLOv5 model is 0.989, slightly inferior to the YOLOv8 network. The mAP comparison between YOLOv8 and the Faster-RCNN network model shows that in this spinal fracture detection task, the performance of the Faster-RCNN network is significantly inferior to that of YOLOv8. This is because the YOLO network excels at detecting small targets. This paper also compares the precision and recall rates of the four models, as shown in Fig. 4 , in addition to mAP. From the graph, it is clear that the trends in precision and recall rates of each model are consistent with the previously described mAP changes. This observation further confirms the outstanding performance of YOLO-v8 in this study. This paper proposes an automated method for detecting fractures caused by cancerous spinal bone metastases using deep learning-based object detection techniques. Four object detection algorithms were employed for detection, including the single-stage algorithms YOLO-v8, YOLO-v7, YOLO-v5, and the two-stage algorithm Faster-RCNN. Through experimental comparison, all four algorithms demonstrated high detection accuracy, with YOLO-v8 achieving 99.1%, YOLO-v7 at 98.0%, YOLO-v5 at 98.9%, and Faster-RCNN at 96.0%. These results indicate that the selected object detection algorithms have all performed well in detecting fractures caused by cancerous spinal bone metastases, particularly the YOLO-v8 network model, which has provided strong technical support for clinical diagnosis with its superior performance. Materials and Methods The data utilized in this study are retrospective and were sourced from the Shengjing Hospital of China Medical University. The dataset encompasses CT imaging data of 140 patients diagnosed with spinal bone metastases and fractures from 2016 to 2023. The research adheres to the principles outlined in the Helsinki Declaration and has been approved by the Institutional Review Board (IRB). All participants provided informed consent. All sensitive information in the CT images has appropriately been anonymized to protect patient privacy. The dataset comprises 77 male and 63 female patients aged 34 to 79. All CT images obtained are in DICOM format, with each patient's data typically consisting of hundreds of CT slices. It is important to note that the number of slices per patient is not fixed due to the variability in diagnostic needs. The slice thickness is 1mm. The dataset for this study is composed of CT cross-sectional images, as illustrated in Fig. 5 . Although the spinal morphological features are visible in these images, it becomes difficult to accurately identify fractures when the spine is in an abnormal state due to the impact of cancer metastases, as shown in Figs. 5 (b) and (c). We transformed the cross-sectional images into sagittal views to facilitate better vertebral fracture detection. The transformed images are presented in Fig. 6 . Figure 6 (a) illustrates the region of vertebra fracture, while Fig. 6 (b) displays a spine without fracture. This image conversion provides convenience for the subsequent training of deep learning models. 4.1 Preprocessing Initially, we tasked two experienced doctors to annotate the fracture locations on the patient's image. When their annotations diverged, we consulted a distinguished radiologist for reconciliation and verification. This study utilized the RadiAnt DICOM Viewer software to convert DICOM files into jpg format images to facilitate the training of deep learning networks. To ensure the efficiency and accuracy of data processing, we implemented batch image processing through Python programming, which enabled the generation of 2D sample images containing essential information, label images, and text files recording positional coordinates in one go. We primarily relied on libraries like Python's 'pydicom' to achieve this batch-processing functionality. To ensure the uniformity of model inputs, we set the dimensions of all output images to 640×640. After processing, we successfully output 2,448 2D images in jpg format, including 1,224 original and 1,224 labeled images. We randomly selected a preprocessed image from the dataset for illustrative purposes, as shown in Fig. 7 . This rigorous data preprocessing workflow has laid a solid foundation for the subsequent training of deep learning models. 4.2 Model Construction Deep learning techniques have achieved remarkable advancements in computer vision, particularly excelling in tasks such as object detection. However, its application in the clinical diagnosis of pathological fractures resulting from spinal bone metastases in cancer is relatively limited. This study employs both single-stage (YOLO) and two-stage (Faster R-CNN) object detection algorithms to enhance the accuracy and efficiency of these complex detection tasks. Through experimental comparative analysis, this study aims to provide a new medical imaging method to assist clinicians in more accurately diagnosing and treating pathological fractures caused by spinal bone metastases from cancer. 4.2.1 YOLO Network Model The YOLO series algorithms have profoundly impacted object detection with their efficient and real-time detection capabilities and continuously optimized detection accuracy, making them a research focus in the domain. This study employed the YOLO-v5, YOLO-v7, and YOLO-v8 network models to detect pathological fractures caused by cancerous spinal bone metastases. YOLO-v8, a standout in single-stage object detection algorithms, shares many similarities in its training process with other similar models, yet it also has its unique features. At the outset of training, images of tumor-induced spinal bone metastases undergo a series of preprocessing steps to meet the input requirements of the network model. The overall architecture of YOLO-v8 is depicted in Fig. 8 . CBS stands for Convolutional Block, which includes SiLU activation and batch normalization among the components. The C2f module can extract fundamental features from the input images through an initial convolutional block, facilitating feature enhancement and fusion. SPPF, which denotes Spatial Pyramid Pooling - Fast, utilizes a streamlined version of spatial pyramid pooling to capture multi-scale information efficiently. 4.2.2 Faster R-CNN Unlike single-stage object detection algorithms, Faster R-CNN employs a two-stage approach for object detection: proposal generation followed by object classification and localization. Faster R-CNN's innovation lies in integrating region proposal generation and object classification into a single network, achieving end-to-end object detection with high accuracy and rapid processing speeds. The network architecture in this study is depicted in Fig. 9 . The feature extraction network is a crucial component of the Faster R-CNN model, primarily responsible for extracting valuable feature information from the input images for subsequent use by the Region Proposal Network (RPN) and the classification networks. Within Faster R-CNN, the feature extraction network initially performs a series of convolutional, activation, and pooling operations on the input image to extract high-level features. These features play a pivotal role in the later stages of the RPN and classification network, aiding the model in more accurately identifying target objects within the image. 4.3 Experimental Parameter Configuration The specific software and hardware parameters for this experiment are shown in Table 2 . Table 2 Configuration of the experimental environment Parameters CPU Intel(R)Core(TM)i7-8700 GPU NVIDIA GeForce GTX 1080Ti Graphics card memory 8G OS Windows10 Deep learning framework PyTorch1.9 Programming language Python3.8 CUDA version 11.3 The configuration of network parameters directly impacts the effectiveness of experimental models. Proper parameter settings can effectively reduce computational burden and significantly enhance model performance. Consistent parameter setup across different network models is critical to fair performance comparison. The relevant parameters used in this experiment are listed in Table 3 . Table 3 Description of experimental parameters Parameter Value Batch size 32 Epoch 180 Learning rate 0.01 Optimizer SGD Weight decay 0.0005 4.4 Model Performance Evaluation This study divided the dataset into a training set and a testing set according to patient cases. Out of 140 patients, a random subset of 14 patients corresponding to 123 images was used as the testing set, while the remaining 126 patients corresponding to 1101 images were used as the training set. The model's prediction performance on the testing set was evaluated using precision, recall, and the mean Average Precision (mAP), with the mAP value as the primary evaluation metric. Declarations Author contributions H.L. and Z.L. were responsible for the overall design of the study; Z.L. was responsible for the interpretation of medical images and the confirmation of annotations; H.L., Y.Z., Z.Q., and N.B. were responsible for the construction, testing, and tuning of the prediction model; H.L., Y.Z. and Z.Q. wrote the main manuscript text; All authors read and approved the final manuscript. Data availability statement The datasets generated during and analyzed during the current study are not publicly available due to hospital regulations but are available from the corresponding author upon reasonable request. Competing Interests Statement The authors declare that they have no competing interests. Funding This work was supported by the Medical and Industrial crossover project, Department of Science and Technology of Liaoning province (No. 2022-YGJC-46). References World Health Organization (2024). Global Cancer Burden growing, Amidst Mounting Need for Services. [online] https://www.who.int/news/item/01-02-2024-global-cancer-burden-growing--amidst-mounting-need-for-services.‌ Tsukamoto, S. et al. (2021). Current Overview of Treatment for Metastatic Bone Disease. Current oncology (Toronto, Ont.) , 28 (5), 3347–3372. https://doi.org/10.3390/curroncol28050290 Wang, J., Fang, Z., Lang, N., Yuan, H., Su, M. Y., & Baldi, P. (2017). A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks. Computers in biology and medicine , 84 , 137–146. https://doi.org/10.1016/j.compbiomed.2017.03.024 Jaipanya, P., & Chanplakorn, P. (2022). Spinal metastasis: narrative reviews of the current evidence and treatment modalities. The Journal of international medical research , 50 (4), 3000605221091665. https://doi.org/10.1177/03000605221091665 Hong, S. H., Chang, B. S., Kim, H., Kang, D. H., & Chang, S. Y. (2022). An Updated Review on the Treatment Strategy for Spinal Metastasis from the Spine Surgeon's Perspective. Asian spine journal , 16 (5), 799–811. https://doi.org/10.31616/asj.2022.0367 Mossa-Basha, M. et al. (2019). Spinal metastasis: diagnosis, management and follow-up. The British journal of radiology , 92 (1103), 20190211. https://doi.org/10.1259/bjr.20190211 Macedo, F., Ladeira, K., Pinho, F., Saraiva, N., Bonito, N., Pinto, L., & Goncalves, F. (2017). Bone Metastases: An Overview. Oncology reviews , 11 (1), 321. https://doi.org/10.4081/oncol.2017.321 Constans, J. P., de Divitiis, E., Donzelli, R., Spaziante, R., Meder, J. F., & Haye, C. (1983). Spinal metastases with neurological manifestations. Review of 600 cases. Journal of neurosurgery , 59 (1), 111–118. https://doi.org/10.3171/jns.1983.59.1.0111 Zhao, Y., Liu, F., & Wang, W. (2023). Treatment progress of spinal metastatic cancer: a powerful tool for improving the quality of life of the patients. Journal of orthopaedic surgery and research , 18 (1), 563. https://doi.org/10.1186/s13018-023-03975-3 Cui, Z., Tian, Y., Feng, H., Yang, Z., Liu, Z. (2019). Unilateral Versus Bilateral Balloon Kyphoplasty for Osteoporotic Vertebral Compression Fractures: A Systematic Review of Overlapping Meta-analyses. Pain physician , 22 (1), 15–28. Barrett-Connor E. (1995). The economic and human costs of osteoporotic fracture. The American journal of medicine , 98 (2A), 3S–8S. https://doi.org/10.1016/s0002-9343(05)80037-3 Tao, H. L., Zhang, H., Jiang, Y. F., Fan, S. S., Wang, H. W., & Zheng, A. T. (2023). The thoracolumbar interfascial block with local anesthesia in osteoporotic vertebral compression fractures treated with percutaneous kyphoplasty provides better analgesia compared with local anesthesia alone: A randomized controlled study. Frontiers in surgery , 10 , 1133637. https://doi.org/10.3389/fsurg.2023.1133637 Gillies, R. J., Kinahan, P. E., & Hricak, H. (2016). Radiomics: Images Are More than Pictures, They Are Data. Radiology , 278 (2), 563–577. https://doi.org/10.1148/radiol.2015151169 Lambin, P. et al. (2017). Radiomics: the bridge between medical imaging and personalized medicine. Nature reviews. Clinical oncology , 14 (12), 749–762. https://doi.org/10.1038/nrclinonc.2017.141 Ahmed, Z., Mohamed, K., Zeeshan, S., & Dong, X. (2020). Artificial intelligence with multi-functional machine learning platform development for better healthcare and precision medicine. Database : the journal of biological databases and curation , 2020 , baaa010. https://doi.org/10.1093/database/baaa010 Li, Y. C., Chen, H. H., Horng-Shing Lu, H., Hondar Wu, H. T., Chang, M. C., & Chou, P. H. (2021). Can a Deep-learning Model for the Automated Detection of Vertebral Fractures Approach the Performance Level of Human Subspecialists?. Clinical orthopaedics and related research , 479 (7), 1598–1612. https://doi.org/10.1097/CORR.0000000000001685 Muehlematter, U. J. et al. (2019). Vertebral body insufficiency fractures: detection of vertebrae at risk on standard CT images using texture analysis and machine learning. European radiology , 29 (5), 2207–2217. https://doi.org/10.1007/s00330-018-5846-8 Oh, E., Seo, S. W., Yoon, Y. C., Kim, D. W., Kwon, S., & Yoon, S. (2017). Prediction of pathologic femoral fractures in patients with lung cancer using machine learning algorithms: Comparison of computed tomography-based radiological features with clinical features versus without clinical features. Journal of orthopaedic surgery (Hong Kong) , 25 (2), 2309499017716243. https://doi.org/10.1177/2309499017716243 Wang, Z., Wen, X., Lu, Y., Yao, Y., & Zhao, H. (2016). Exploiting machine learning for predicting skeletal-related events in cancer patients with bone metastases. Oncotarget , 7(11) , 12612–12622. https://doi.org/10.18632/oncotarget.7278 Ramachandran, S., George, J., Skaria, S., & Varun, V. V. (2018). Using YOLO based deep learning network for real time detection and localization of lung nodules from low dose CT scans. [C]//Medical Imaging 2018: Computer-Aided Diagnosis. SPIE, 2018, 10575 : 347-355. Xu, L. et al. (2017). W-Net for Whole-Body Bone Lesion Detection on 68 Ga-Pentixafor PET/CT Imaging of Multiple Myeloma Patients. https://doi.org/10.1007/978-3-319-67564-0_3. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5364269","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":388469878,"identity":"97f145e8-b1b4-4e0e-a9a4-cb0ad3c80f5d","order_by":0,"name":"Yaowen Zhang","email":"","orcid":"","institution":"Northeastern University","correspondingAuthor":false,"prefix":"","firstName":"Yaowen","middleName":"","lastName":"Zhang","suffix":""},{"id":388469879,"identity":"29a1b429-9fc9-491f-b3c8-7b3b47c99849","order_by":1,"name":"Zihan Qin","email":"","orcid":"","institution":"Northeastern University","correspondingAuthor":false,"prefix":"","firstName":"Zihan","middleName":"","lastName":"Qin","suffix":""},{"id":388469880,"identity":"981a1260-29b5-4de6-98ab-7106fb2c1410","order_by":2,"name":"Nan Bao","email":"","orcid":"","institution":"Northeastern University","correspondingAuthor":false,"prefix":"","firstName":"Nan","middleName":"","lastName":"Bao","suffix":""},{"id":388469881,"identity":"6c2c5805-68fe-44a9-a347-d9155bf1dc2d","order_by":3,"name":"Hong Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIiWNgGAWjYBACPmYeEGWTwMAMF0vAr4UNoiWNFC0MYC2HkZUR0sLOe/Bxwa/zebrtzAeYCyoOM/Cz5xgw/NyBz2F8ycYz+24Xmx1mS2CeceYwg2TPGwPG3jN4/WImzdtzO3HbYR4DZt62wwwGN3IMmBnbCGo5B9TC/4GZ999hBnuitPD8OACyhYGZtwFoiwRhLcbGvA3JIL8YHOY5ls4jceZZwcFePFr4+c8YPub5Y5dndv7ww8c8NdZy/O3JGx/8xKMFDGDOOADEPDAGAfCHsJJRMApGwSgYwQAAimFG+UPzPbIAAAAASUVORK5CYII=","orcid":"","institution":"Northeastern University","correspondingAuthor":true,"prefix":"","firstName":"Hong","middleName":"","lastName":"Li","suffix":""},{"id":388469882,"identity":"81a07b96-ae44-4172-8e90-35b94d5b2e08","order_by":4,"name":"Zhen Liu","email":"","orcid":"","institution":"Shengjing Hospital of China Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhen","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2024-10-31 02:38:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5364269/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5364269/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":71688259,"identity":"a069db4b-ee9d-4350-913e-eb223ec3745d","added_by":"auto","created_at":"2024-12-17 17:39:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":156344,"visible":true,"origin":"","legend":"\u003cp\u003eThe training-loss curves correspond to the four models.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/074a10836bf7a1fa63a13c6e.png"},{"id":71688257,"identity":"a9bef6ed-a584-4fef-93bc-28dae779bf0a","added_by":"auto","created_at":"2024-12-17 17:39:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":378374,"visible":true,"origin":"","legend":"\u003cp\u003ePrediction Results of YOLO-v8.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/fd5ffc28deecb4f8997bfdb2.png"},{"id":71688258,"identity":"7ac69b11-3a0f-4864-944b-a765d7c74160","added_by":"auto","created_at":"2024-12-17 17:39:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":95826,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of mAP between YOLO-v8 and other networks.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/a2d776e7ea037c73126918db.png"},{"id":71688956,"identity":"62bbac55-8897-4824-b608-78d2c177d634","added_by":"auto","created_at":"2024-12-17 17:47:25","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":127546,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of different models. (a) Precision; (b) Recall.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/b44eede3843238a0889202b6.png"},{"id":71688264,"identity":"e2950381-ca92-4b38-b97c-2ec74694d12e","added_by":"auto","created_at":"2024-12-17 17:39:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":180463,"visible":true,"origin":"","legend":"\u003cp\u003eCT Cross-Sectional Images. (a) Normal vertebrae; (b) Vertebrae with cancer bone metastases but without fracture; (c)Vertebrae fracture with cancer bone metastases.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/18c1040803261eb45bd35895.png"},{"id":71688953,"identity":"23145bf9-1c7a-4564-a2e6-af1d45caf0be","added_by":"auto","created_at":"2024-12-17 17:47:24","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":126369,"visible":true,"origin":"","legend":"\u003cp\u003eCT Sagittal Images, with the positions of the horizontal lines corresponding to the cross-sectional locations shown in Figures 5(b) and (c). (a) Fracture at the position of the horizontal line; (b) No fracture at the horizontal line position.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/cf9e06388d918c761e6de166.png"},{"id":71689918,"identity":"8a63e7f4-842d-44f6-8bf7-e9a2acead7ed","added_by":"auto","created_at":"2024-12-17 17:55:25","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":135075,"visible":true,"origin":"","legend":"\u003cp\u003ePreprocessed Sagittal View Image and its Label Image (a) Sagittal View Image (b) Fracture Label Image.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/d8393800ad8126a74ba2eef5.png"},{"id":71688955,"identity":"950594fa-165f-448a-9937-a9df48b75a5f","added_by":"auto","created_at":"2024-12-17 17:47:24","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":81079,"visible":true,"origin":"","legend":"\u003cp\u003eThe Architecture of the YOLO-v8 Model Used in This Study.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/6f7dc49e7eb0f2b60f4e0a41.png"},{"id":71688262,"identity":"8afce1bf-b90d-4ff5-8287-0083e704159d","added_by":"auto","created_at":"2024-12-17 17:39:24","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":53876,"visible":true,"origin":"","legend":"\u003cp\u003eTwo-Stage Detection of Faster R-CNN in This Study.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/fa9c6b908e430e6617563660.png"},{"id":71764080,"identity":"1347198d-5b6c-4370-8dee-60be157afb60","added_by":"auto","created_at":"2024-12-18 11:23:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1773582,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5364269/v1/5d9ba171-b41d-4ebc-bbe8-b6088991d1dc.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Comparative Study of AI Models for Pathological Fractures Identification in Spinal Metastases","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAccording to data statistics from the World Health Organization (WHO), approximately 20\u0026nbsp;million new cancer cases were reported globally in 2022, with nearly 10\u0026nbsp;million cancer-related deaths [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Bone metastasis is a common complication in most malignant tumors, with the spine being a frequent site of metastasis [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Statistics show that about 70% of cancer patients will develop spinal metastases [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], and 10%-20% of these cases will experience concurrent symptoms, leading to pain, neurological deficits, and a deterioration in quality of life [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Moreover, 10% of patients may suffer from spinal cord compression, which can result in potentially permanent disability [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe bone destruction caused by spinal metastases mainly manifests in three types: osteoblastic, lytic, and mixed [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Data indicate that lytic spinal metastases occur more frequently, accounting for about 70% of cases [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. The incidence of pathological fractures due to bone metastases ranges from 10\u0026ndash;29% [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. As tumors grow and affect the vertebral body or compress the spinal cord and nerve roots, symptoms such as pain, bowel dysfunction, sensory and motor dysfunction of the lower limbs, and even paralysis may arise [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Therefore, early detection and prediction of pathological fractures in spinal metastases are crucial for improving the quality of life of patients with advanced cancer [\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn recent years, radiomics technology has rapidly developed, which uses computer analysis to extract features from medical images, transforming them into high-dimensional data for analysis [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. This technology has the potential to improve disease diagnosis and prognosis. Deep learning can autonomously learn data features as part of machine learning, reducing manual processing. Integrating deep learning with machine learning and radiomics has shown significance in cancer research [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Muehlematter et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] employed machine learning and texture analysis techniques to evaluate consecutive CT scan images of 58 patients with spinal insufficiency fractures in spinal fracture detection. The study found that a Support Vector Machine (SVM) classifier combined with 29 texture parameter features achieved a high AUC (Area Under Curve) of 0.97 in identifying patients with spinal insufficiency. Oh et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] studied 84 lung cancer patients with proximal femoral bone metastases who underwent CT scans and were followed up for at least three months. They extracted and standardized 16 clinical and radiological features and then trained these features using five supervised learning classifiers. The study revealed that the Gradient Boosting (GB) algorithm had the best predictive performance, with an AUC of 0.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.14.\u003c/p\u003e \u003cp\u003eAdditionally, Wang et al. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] collected clinical and radiological data from 1143 patients with bone metastases and used machine learning techniques to predict Skeletal-Related Events (SREs). They found that an SVM classifier with eight variables achieved an accuracy of 97.1% in predicting SREs. Ramachandran et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] utilized a single-stage YOLO (You Only Look Once) network structure based on the DetectNet architecture for detecting pulmonary nodules in lung CT images. This method achieved a meager false-positive rate, demonstrating high sensitivity and accuracy. Xu et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] designed an innovative cascaded convolutional neural network, W-Net, to learn and extract critical features from multi-modal medical images autonomously. These research outcomes suggest that deep learning-based object detection technologies have potential applications in medical diagnostics.\u003c/p\u003e \u003cp\u003eTherefore, this study aims to explore the use of deep learning networks for the target detection of pathological fractures caused by cancerous spinal bone metastases.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eIn this study, models such as Faster R-CNN, YOLO-v5, YOLO-v7, and YOLO-v8 were constructed on the training dataset. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the training-loss curves for the Faster R-CNN, YOLO-v5, YOLO-v7, and YOLO-v8 models. It can be observed that all four curves decrease over iterations and stabilize, demonstrating that the models were trained effectively. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the prediction results of YOLO-v8, where the top row is the label images and the bottom row is the prediction results. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the precision, recall, and mAP corresponding to the prediction results of the four models on the test set.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Prediction performance on the testing set.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003emAP\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFasterR-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.952\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.964\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.960\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO-v5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.983\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.990\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.989\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO-v7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.977\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.980\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYOLO-v8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.999\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.998\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.991\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study utilized four object detection models to detect fractures on CT images of cancer patients with spinal bone metastases. The experimental results indicate that all four models achieved satisfactory detection after thorough learning, training, and appropriate parameter adjustments.\u003c/p\u003e \u003cp\u003eIf mAP is used as the critical evaluation index, we can observe that YOLO-v8 has a clear advantage over other networks. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the comparison results between YOLO-v8 and YOLO-v7, YOLO-v5, and Faster R-CNN.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe mAP comparison between YOLOv8 and YOLOv7 shows that the former has a significantly better convergence speed than the latter. From the initial few epochs, the performance of the YOLOv8 model starts to outperform YOLOv7 significantly, and the mAP value of YOLOv8 reaches 0.991, slightly higher than that of the comparative network.\u003c/p\u003e \u003cp\u003eIn the comparison between YOLOv8 and YOLOv5, it can be seen from the graph that the mAP curves of the two models generally remain consistent. Initially, the two curves change more or less in unison, but at around 30 epochs, the performance of the YOLOv8 model begins to surpass YOLOv5. In the end, the mAP value of the YOLOv5 model is 0.989, slightly inferior to the YOLOv8 network.\u003c/p\u003e \u003cp\u003eThe mAP comparison between YOLOv8 and the Faster-RCNN network model shows that in this spinal fracture detection task, the performance of the Faster-RCNN network is significantly inferior to that of YOLOv8. This is because the YOLO network excels at detecting small targets.\u003c/p\u003e \u003cp\u003eThis paper also compares the precision and recall rates of the four models, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, in addition to mAP. From the graph, it is clear that the trends in precision and recall rates of each model are consistent with the previously described mAP changes. This observation further confirms the outstanding performance of YOLO-v8 in this study.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThis paper proposes an automated method for detecting fractures caused by cancerous spinal bone metastases using deep learning-based object detection techniques. Four object detection algorithms were employed for detection, including the single-stage algorithms YOLO-v8, YOLO-v7, YOLO-v5, and the two-stage algorithm Faster-RCNN. Through experimental comparison, all four algorithms demonstrated high detection accuracy, with YOLO-v8 achieving 99.1%, YOLO-v7 at 98.0%, YOLO-v5 at 98.9%, and Faster-RCNN at 96.0%. These results indicate that the selected object detection algorithms have all performed well in detecting fractures caused by cancerous spinal bone metastases, particularly the YOLO-v8 network model, which has provided strong technical support for clinical diagnosis with its superior performance.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eThe data utilized in this study are retrospective and were sourced from the Shengjing Hospital of China Medical University. The dataset encompasses CT imaging data of 140 patients diagnosed with spinal bone metastases and fractures from 2016 to 2023. The research adheres to the principles outlined in the Helsinki Declaration and has been approved by the Institutional Review Board (IRB). All participants provided informed consent. All sensitive information in the CT images has appropriately been anonymized to protect patient privacy. The dataset comprises 77 male and 63 female patients aged 34 to 79. All CT images obtained are in DICOM format, with each patient\u0026apos;s data typically consisting of hundreds of CT slices. It is important to note that the number of slices per patient is not fixed due to the variability in diagnostic needs. The slice thickness is 1mm.\u003c/p\u003e\n\u003cp\u003eThe dataset for this study is composed of CT cross-sectional images, as illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. Although the spinal morphological features are visible in these images, it becomes difficult to accurately identify fractures when the spine is in an abnormal state due to the impact of cancer metastases, as shown in Figs. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e(b) and (c).\u003c/p\u003e\n\u003cp\u003eWe transformed the cross-sectional images into sagittal views to facilitate better vertebral fracture detection. The transformed images are presented in Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e. Figure \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e(a) illustrates the region of vertebra fracture, while Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e(b) displays a spine without fracture. This image conversion provides convenience for the subsequent training of deep learning models.\u003c/p\u003e\n\u003ch3\u003e4.1 Preprocessing\u003c/h3\u003e\n\u003cp\u003eInitially, we tasked two experienced doctors to annotate the fracture locations on the patient\u0026apos;s image. When their annotations diverged, we consulted a distinguished radiologist for reconciliation and verification. This study utilized the RadiAnt DICOM Viewer software to convert DICOM files into jpg format images to facilitate the training of deep learning networks. To ensure the efficiency and accuracy of data processing, we implemented batch image processing through Python programming, which enabled the generation of 2D sample images containing essential information, label images, and text files recording positional coordinates in one go. We primarily relied on libraries like Python\u0026apos;s \u0026apos;pydicom\u0026apos; to achieve this batch-processing functionality. To ensure the uniformity of model inputs, we set the dimensions of all output images to 640\u0026times;640. After processing, we successfully output 2,448 2D images in jpg format, including 1,224 original and 1,224 labeled images. We randomly selected a preprocessed image from the dataset for illustrative purposes, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e. This rigorous data preprocessing workflow has laid a solid foundation for the subsequent training of deep learning models.\u003c/p\u003e\n\u003ch3\u003e4.2 Model Construction\u003c/h3\u003e\n\u003cp\u003eDeep learning techniques have achieved remarkable advancements in computer vision, particularly excelling in tasks such as object detection. However, its application in the clinical diagnosis of pathological fractures resulting from spinal bone metastases in cancer is relatively limited. This study employs both single-stage (YOLO) and two-stage (Faster R-CNN) object detection algorithms to enhance the accuracy and efficiency of these complex detection tasks. Through experimental comparative analysis, this study aims to provide a new medical imaging method to assist clinicians in more accurately diagnosing and treating pathological fractures caused by spinal bone metastases from cancer.\u003c/p\u003e\n\u003ch3\u003e4.2.1 YOLO Network Model\u003c/h3\u003e\n\u003cp\u003eThe YOLO series algorithms have profoundly impacted object detection with their efficient and real-time detection capabilities and continuously optimized detection accuracy, making them a research focus in the domain. This study employed the YOLO-v5, YOLO-v7, and YOLO-v8 network models to detect pathological fractures caused by cancerous spinal bone metastases. YOLO-v8, a standout in single-stage object detection algorithms, shares many similarities in its training process with other similar models, yet it also has its unique features. At the outset of training, images of tumor-induced spinal bone metastases undergo a series of preprocessing steps to meet the input requirements of the network model. The overall architecture of YOLO-v8 is depicted in Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e. CBS stands for Convolutional Block, which includes SiLU activation and batch normalization among the components. The C2f module can extract fundamental features from the input images through an initial convolutional block, facilitating feature enhancement and fusion. SPPF, which denotes Spatial Pyramid Pooling - Fast, utilizes a streamlined version of spatial pyramid pooling to capture multi-scale information efficiently.\u003c/p\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2.2 Faster R-CNN\u003c/h2\u003e\n \u003cp\u003eUnlike single-stage object detection algorithms, Faster R-CNN employs a two-stage approach for object detection: proposal generation followed by object classification and localization. Faster R-CNN\u0026apos;s innovation lies in integrating region proposal generation and object classification into a single network, achieving end-to-end object detection with high accuracy and rapid processing speeds. The network architecture in this study is depicted in Fig. \u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e. The feature extraction network is a crucial component of the Faster R-CNN model, primarily responsible for extracting valuable feature information from the input images for subsequent use by the Region Proposal Network (RPN) and the classification networks. Within Faster R-CNN, the feature extraction network initially performs a series of convolutional, activation, and pooling operations on the input image to extract high-level features. These features play a pivotal role in the later stages of the RPN and classification network, aiding the model in more accurately identifying target objects within the image.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003e4.3 Experimental Parameter Configuration\u003c/h3\u003e\n\u003cp\u003eThe specific software and hardware parameters for this experiment are shown in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eConfiguration of the experimental environment\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameters\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIntel(R)Core(TM)i7-8700\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNVIDIA GeForce GTX 1080Ti\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGraphics card memory\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8G\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWindows10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDeep learning framework\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePyTorch1.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProgramming language\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePython3.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCUDA version\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe configuration of network parameters directly impacts the effectiveness of experimental models. Proper parameter settings can effectively reduce computational burden and significantly enhance model performance. Consistent parameter setup across different network models is critical to fair performance comparison. The relevant parameters used in this experiment are listed in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDescription of experimental parameters\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameter\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eValue\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBatch size\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEpoch\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e180\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLearning rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSGD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWeight decay\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.0005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003ch3\u003e4.4 Model Performance Evaluation\u003c/h3\u003e\n\u003cp\u003eThis study divided the dataset into a training set and a testing set according to patient cases. Out of 140 patients, a random subset of 14 patients corresponding to 123 images was used as the testing set, while the remaining 126 patients corresponding to 1101 images were used as the training set. The model\u0026apos;s prediction performance on the testing set was evaluated using precision, recall, and the mean Average Precision (mAP), with the mAP value as the primary evaluation metric.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eH.L. and Z.L. were responsible for the overall design of the study; Z.L. was responsible for the interpretation of medical images and the confirmation of annotations; H.L., Y.Z., Z.Q., and N.B. were responsible for the construction, testing, and tuning of the prediction model; H.L., Y.Z. and Z.Q. wrote the main manuscript text; All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated during and analyzed during the current study are not publicly available due to hospital regulations but are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Medical and Industrial crossover project, Department of Science and Technology of Liaoning province (No. 2022-YGJC-46).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eWorld Health Organization (2024). Global Cancer Burden growing, Amidst Mounting Need for Services. [online] https://www.who.int/news/item/01-02-2024-global-cancer-burden-growing--amidst-mounting-need-for-services.\u0026zwnj;\u003c/li\u003e\n\u003cli\u003eTsukamoto, S. et al. (2021). Current Overview of Treatment for Metastatic Bone Disease. \u003cem\u003eCurrent oncology (Toronto, Ont.)\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(5), 3347\u0026ndash;3372. https://doi.org/10.3390/curroncol28050290\u003c/li\u003e\n\u003cli\u003eWang, J., Fang, Z., Lang, N., Yuan, H., Su, M. Y., \u0026amp; Baldi, P. (2017). A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks. \u003cem\u003eComputers in biology and medicine\u003c/em\u003e, \u003cem\u003e84\u003c/em\u003e, 137\u0026ndash;146. https://doi.org/10.1016/j.compbiomed.2017.03.024\u003c/li\u003e\n\u003cli\u003eJaipanya, P., \u0026amp; Chanplakorn, P. (2022). Spinal metastasis: narrative reviews of the current evidence and treatment modalities. \u003cem\u003eThe Journal of international medical research\u003c/em\u003e, \u003cem\u003e50\u003c/em\u003e(4), 3000605221091665. https://doi.org/10.1177/03000605221091665\u003c/li\u003e\n\u003cli\u003eHong, S. H., Chang, B. S., Kim, H., Kang, D. H., \u0026amp; Chang, S. Y. (2022). An Updated Review on the Treatment Strategy for Spinal Metastasis from the Spine Surgeon\u0026apos;s Perspective. \u003cem\u003eAsian spine journal\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(5), 799\u0026ndash;811. https://doi.org/10.31616/asj.2022.0367\u003c/li\u003e\n\u003cli\u003eMossa-Basha, M. et al. (2019). Spinal metastasis: diagnosis, management and follow-up. \u003cem\u003eThe British journal of radiology\u003c/em\u003e, \u003cem\u003e92\u003c/em\u003e(1103), 20190211. https://doi.org/10.1259/bjr.20190211\u003c/li\u003e\n\u003cli\u003eMacedo, F., Ladeira, K., Pinho, F., Saraiva, N., Bonito, N., Pinto, L., \u0026amp; Goncalves, F. (2017). Bone Metastases: An Overview. \u003cem\u003eOncology reviews\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(1), 321. https://doi.org/10.4081/oncol.2017.321 \u003c/li\u003e\n\u003cli\u003eConstans, J. P., de Divitiis, E., Donzelli, R., Spaziante, R., Meder, J. F., \u0026amp; Haye, C. (1983). Spinal metastases with neurological manifestations. Review of 600 cases. \u003cem\u003eJournal of neurosurgery\u003c/em\u003e, \u003cem\u003e59\u003c/em\u003e(1), 111\u0026ndash;118. https://doi.org/10.3171/jns.1983.59.1.0111 \u003c/li\u003e\n\u003cli\u003eZhao, Y., Liu, F., \u0026amp; Wang, W. (2023). Treatment progress of spinal metastatic cancer: a powerful tool for improving the quality of life of the patients. \u003cem\u003eJournal of orthopaedic surgery and research\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e(1), 563. https://doi.org/10.1186/s13018-023-03975-3\u003c/li\u003e\n\u003cli\u003eCui, Z., Tian, Y., Feng, H., Yang, Z., Liu, Z. (2019). Unilateral Versus Bilateral Balloon Kyphoplasty for Osteoporotic Vertebral Compression Fractures: A Systematic Review of Overlapping Meta-analyses. \u003cem\u003ePain physician\u003c/em\u003e, \u003cem\u003e22\u003c/em\u003e(1), 15\u0026ndash;28.\u003c/li\u003e\n\u003cli\u003eBarrett-Connor E. (1995). The economic and human costs of osteoporotic fracture. \u003cem\u003eThe American journal of medicine\u003c/em\u003e, \u003cem\u003e98\u003c/em\u003e(2A), 3S\u0026ndash;8S. https://doi.org/10.1016/s0002-9343(05)80037-3 \u003c/li\u003e\n\u003cli\u003eTao, H. L., Zhang, H., Jiang, Y. F., Fan, S. S., Wang, H. W., \u0026amp; Zheng, A. T. (2023). The thoracolumbar interfascial block with local anesthesia in osteoporotic vertebral compression fractures treated with percutaneous kyphoplasty provides better analgesia compared with local anesthesia alone: A randomized controlled study. \u003cem\u003eFrontiers in surgery\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 1133637. https://doi.org/10.3389/fsurg.2023.1133637 \u003c/li\u003e\n\u003cli\u003eGillies, R. J., Kinahan, P. E., \u0026amp; Hricak, H. (2016). Radiomics: Images Are More than Pictures, They Are Data. \u003cem\u003eRadiology\u003c/em\u003e, \u003cem\u003e278\u003c/em\u003e(2), 563\u0026ndash;577. https://doi.org/10.1148/radiol.2015151169 \u003c/li\u003e\n\u003cli\u003eLambin, P. et al. (2017). Radiomics: the bridge between medical imaging and personalized medicine. \u003cem\u003eNature reviews. Clinical oncology\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(12), 749\u0026ndash;762. https://doi.org/10.1038/nrclinonc.2017.141 \u003c/li\u003e\n\u003cli\u003eAhmed, Z., Mohamed, K., Zeeshan, S., \u0026amp; Dong, X. (2020). Artificial intelligence with multi-functional machine learning platform development for better healthcare and precision medicine. \u003cem\u003eDatabase : the journal of biological databases and curation\u003c/em\u003e, \u003cem\u003e2020\u003c/em\u003e, baaa010. https://doi.org/10.1093/database/baaa010\u003c/li\u003e\n\u003cli\u003eLi, Y. C., Chen, H. H., Horng-Shing Lu, H., Hondar Wu, H. T., Chang, M. C., \u0026amp; Chou, P. H. (2021). Can a Deep-learning Model for the Automated Detection of Vertebral Fractures Approach the Performance Level of Human Subspecialists?. \u003cem\u003eClinical orthopaedics and related research\u003c/em\u003e, \u003cem\u003e479\u003c/em\u003e(7), 1598\u0026ndash;1612. https://doi.org/10.1097/CORR.0000000000001685\u003c/li\u003e\n\u003cli\u003eMuehlematter, U. J. et al. (2019). Vertebral body insufficiency fractures: detection of vertebrae at risk on standard CT images using texture analysis and machine learning. \u003cem\u003eEuropean radiology\u003c/em\u003e, \u003cem\u003e29\u003c/em\u003e(5), 2207\u0026ndash;2217. https://doi.org/10.1007/s00330-018-5846-8\u003c/li\u003e\n\u003cli\u003eOh, E., Seo, S. W., Yoon, Y. C., Kim, D. W., Kwon, S., \u0026amp; Yoon, S. (2017). Prediction of pathologic femoral fractures in patients with lung cancer using machine learning algorithms: Comparison of computed tomography-based radiological features with clinical features versus without clinical features. \u003cem\u003eJournal of orthopaedic surgery (Hong Kong)\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e(2), 2309499017716243. https://doi.org/10.1177/2309499017716243\u003c/li\u003e\n\u003cli\u003eWang, Z., Wen, X., Lu, Y., Yao, Y., \u0026amp; Zhao, H. (2016). Exploiting machine learning for predicting skeletal-related events in cancer patients with bone metastases. \u003cem\u003eOncotarget\u003c/em\u003e, \u003cem\u003e7(11)\u003c/em\u003e, 12612\u0026ndash;12622. https://doi.org/10.18632/oncotarget.7278\u003c/li\u003e\n\u003cli\u003eRamachandran, S., George, J., Skaria, S., \u0026amp; Varun, V. V. (2018). Using YOLO based deep learning network for real time detection and localization of lung nodules from low dose CT scans. \u003cem\u003e[C]//Medical Imaging 2018: Computer-Aided Diagnosis. SPIE, 2018, 10575\u003c/em\u003e: 347-355.\u003c/li\u003e\n\u003cli\u003eXu, L. et al. (2017). W-Net for Whole-Body Bone Lesion Detection on \u003csup\u003e68\u003c/sup\u003eGa-Pentixafor PET/CT Imaging of Multiple Myeloma Patients. https://doi.org/10.1007/978-3-319-67564-0_3.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"vertebral fracture, fracture identification, spinal cancer metastases, deep learning, YOLO","lastPublishedDoi":"10.21203/rs.3.rs-5364269/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5364269/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBone metastasis is a common complication in most malignant tumors, with the spine being a frequent site of metastasis. Cancerous spinal bone metastases can easily lead to pathological fractures, resulting in pain, neurological deficits, and a deterioration in the quality of life for patients. Therefore, early identification of pathological fractures in spinal metastases is crucial for improving the survival quality of patients with advanced cancer. This retrospective study collected CT imaging data from 140 patients with spinal bone metastases and resultant fractures, including 77 male and 63 female patients, with an age range of 34 to 79 years. The study began by preprocessing the CT images and had experienced radiologists annotate the fractures. Subsequently, deep learning-based models, single-stage YOLO-v5, YOLO-v7, and YOLO-v8, as well as the two-stage Faster R-CNN, were constructed and analyzed to detect pathological fractures caused by cancerous spinal metastases. Experimental results indicate that these AI models achieved favorable detection performance, with YOLO-v8 demonstrating the best performance on an independent testing set, achieving a precision of 0.999, a recall rate of 0.998, and a mean average precision (mAP) of 0.991. The vertebral\u003cstrong\u003e \u003c/strong\u003efracture identification model based on YOLO-v8 can provide strong technical support for clinical diagnosis.\u003c/p\u003e","manuscriptTitle":"Comparative Study of AI Models for Pathological Fractures Identification in Spinal Metastases","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-17 17:39:20","doi":"10.21203/rs.3.rs-5364269/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"6eb992dd-dda4-4271-b8fd-adc646caa434","owner":[],"postedDate":"December 17th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":41397596,"name":"Health sciences/Health care/Diagnosis"},{"id":41397597,"name":"Health sciences/Health care/Medical imaging/Bone imaging"}],"tags":[],"updatedAt":"2024-12-18T11:23:10+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-17 17:39:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5364269","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5364269","identity":"rs-5364269","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}