Postoperative One Year Prediction for Patients with Cervical Spinal Cord Injury Based on Deep Learning and Radiomics

doi:10.21203/rs.3.rs-4848654/v1

Download PDF

Article

Postoperative One Year Prediction for Patients with Cervical Spinal Cord Injury Based on Deep Learning and Radiomics

https://doi.org/10.21203/rs.3.rs-4848654/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background:

Cervical spinal cord injury (SCI) can lead to significant impairments, requiring extensive care and posing considerable challenges in predicting postoperative outcomes. This study aimed to develop and validate a deep learning radiomics (DLR) model combining deep learning and radiomics features to improve the prognostic prediction of cervical SCI.

Methods:

This retrospective study included 82 patients with confirmed cervical SCI from three hospitals, collected between January 2012 and January 2021. Patients were divided into good prognosis and poor prognosis groups based on postoperative ASIA grade improvement. Preoperative MRI images were processed using various filtering techniques, and regions of interest (ROI) were segmented and analyzed to extract radiomics features. Deep learning models (ResNet-18, ResNet-50, and ResNet-101) were trained. Features from both radiomics and deep learning models were combined and selected 、 to build the final predictive model using MLP.

Results:

ResNet-50 outperformed other models, demonstrating an AUC of 0.8750 in the test set. The combined model (Rad + ResNet-50) showed the highest prognostic value with an AUC of 0.9220 in the test set. Grad-CAM images enhanced the interpretability of the model by highlighting critical areas for prognosis prediction.

Conclusion:

Integrating deep learning and radiomics features significantly improves the prediction accuracy for cervical SCI outcomes. The Rad + ResNet-50 model, with its superior performance and interpretability, holds promise for clinical applications, offering a robust tool for predicting functional prognosis in cervical SCI patients. Further prospective studies with larger datasets are needed to validate these findings.

Health sciences/Neurology/Neurological disorders/Neuromuscular disease

Biological sciences/Neuroscience/Diseases of the nervous system/Neurodegeneration

Cervical spinal cord injury (SCI) can result in a spectrum of outcomes, from complete to incomplete tetraplegia, leading to significant physical, psychological, and social consequences for those affected [1]. Individuals with cervical SCI experience substantial impairments in various aspects of life, requiring extensive care and effort. Preoperative prediction is crucial for postoperative patients, yet the associated risk factors remain unidentified [2, 3].

The convergence of machine learning (ML) and medical advancements has significantly improved diagnosis, risk assessment, and treatment response prediction [4–7]. ML algorithms can identify subtle data patterns often missed by human experts, though research on cervical SCI remains limited [8–11]. Radiomics, which extracts numerous quantitative features from medical images and analyzes them using advanced algorithms, has been employed to detect spinal cord compression at different locations and guide management decisions [12].

Deep learning (DL), utilizing convolutional neural networks to extract image features, has driven significant advancements in artificial intelligence and human-computer interactions. This approach reduces diagnostic errors and maximizes the potential of radiomics data. Furthermore, deep transfer learning enables the application of deep learning to small datasets. As neuroscience advances intersect with technological innovations, neurosurgery finds unique opportunities to optimize patient care through ANNs. However, integrating clinical relevance with complex, non-linear data for accurate prognostic modeling remains challenging[12].

This study aimed to compare traditional radiomics with deep learning-enhanced radiomics and to develop and validate a deep learning radiomics (DLR) model for accurately predicting cervical SCI outcomes.

Patient Selection and Study Design

This retrospective study received approval from the ethical committees of three hospitals (approval no. 2021KY138), with informed consent waived. Data from 223 patients collected between January 2012 and January 2021 were included.

Inclusion criteria:

Confirmed diagnosis of cervical spinal cord injury (acute injury to nerve structures in the spinal canal causing temporary or permanent sensory and motor function changes, with or without vesicorectal dysfunction)

Complete case data and preoperative MRI images

No prior treatment before surgery

Exclusion criteria

Incomplete clinical records or inadequate imaging data.

Initially, 223 patients were included: 118 from Ordos Central Hospital, and 19 from The First People’s Hospital of ChangDe City, 86 from Fujian Medical University Union Hospital. We excluded 86 patients with non-cervical spinal cord injuries. Of the remaining 137 patients, 55 were excluded due to incomplete or low-quality MRI images. Ultimately, 82 patients were analyzed, divided into good prognosis (41 patients) and poor prognosis (41 patients). The patient recruitment map is shown in Fig. 1A, and the study design and pipeline are illustrated in Fig. 1C.

Surgical Procedures

All participants underwent neck surgery via either an anterior or posterior approach, including anterior cervical discectomy, microdecompression of the spinal canal, implant fusion, internal fixation with steel plate, or posterior median cervical internal fixation, implant fusion, and decompression. Based on the study by Yanlin Yin et al., there were no significant differences in postoperative cervical and neurological recovery between anterior and posterior surgical interventions[13]. Follow-ups were conducted to monitor progress. Detailed descriptions of the surgical techniques are provided below:

Patient Position: Prone for the posterior approach, supine for the anterior approach.

Anesthesia Method: Tracheal intubation with sedation combined with general anesthesia.

Preoperative Localization: Head frame fixation after retraction (posterior approach) and immobilization. Using X-ray front and side fluoroscopy of the C-arm machine, the surgical segments were localized and marked.

Localization of Puncture: After routine disinfection and towel spreading, the surgical incision was made according to intraoperative C-arm X-ray positive and lateral fluoroscopy film, with layer-by-layer incision of the skin, subcutaneous tissue, and deep fascia.

Surgical Steps[13, 14]:

Anterior Approach: Vertebral body resection and thorough excision of implicated intervertebral discs. Cervical fusion was achieved using either allograft fibular struts or autografts, secured with semi-rigid locking plates. Postoperatively, cervical immobilization with a collar was maintained for two weeks until wound healing.
Posterior Approach: Using a Mayfield clamp or horseshoe headrest, the cervical spine was stabilized in a neutral position. The extent of cervical laminoplasty was predetermined through preoperative imaging. Transverse mass screws were deployed and affixed with connecting rods to achieve segmental decompression by excising the ligamenta flava and laminae of the targeted segments.

Postoperative Care: Complete hemostasis was ensured, paravertebral muscles were repositioned, and the subcutaneous layer and skin were sutured. The operative site was dressed with sterile gauze, and a drainage tube was inserted. Patients were encouraged to start range-of-motion exercises immediately after the drainage tube was removed and were discharged once they were able to walk. Both cohorts were managed in alignment with the hospital’s clinical guidelines for cervical spine interventions, incorporating a uniform regimen of physical therapy.

Clinical Data Collection and ASIA Scale Assessment

Patient demographic and clinical information, including age, gender, and cause of spinal cord injury, were collected. The preoperative and postoperative ASIA scales were utilized to calculate the difference. During postoperative follow-up, the ASIA score criteria remained consistent with preoperative standards. Patients were classified into good prognosis and poor prognosis groups based on whether their ASIA grade improved by at least one level post-surgery: a decrease indicated a good prognosis, otherwise poor prognosis. Regarding the ASIA grade scoring criteria, please refer to Supplementary File 1.

Acquisition and Preprocessing of T2-weighted MRI Images

The preoperative MRI scans for all enrolled patients, featuring T2-weighted imaging (T2WI), were conducted using 3.0 Tesla MRI technology. The machine settings at different centers are shown in Table 3. After obtaining the T2-weighted MRI images, preprocessing included normalization to the range of 0–1, resampling to a voxel size of 1x1x1, and processing image noise using multiple filters: Median, Average Gaussian and Box filtering. Each filter outputted as a separate channel rather than being overlaid. These operations aimed at reducing multicenter effects and enhancing the model's generalization ability. Detailed descriptions of each filter can be found in the supplementary file2.

ROI Segmentation and Radiomics Analysis

Segmentation of Regions of Interest (ROI)

The 3D Slicer Software Application v5.0.2 (https://www.slicer.org/) was used to segment the Region of Interest (ROI). The ROI comprised two segments: the spinal cord injury region (segmentation 1) and the unaffected spinal cord regions (segmentation 2). Segmentation 1 is defined as the area of evident spinal cord injury, such as edema and significant spinal cord compression [15–17]. Segmentation 2 is defined as the region excluding the anatomical landmarks of Segmentation 1, including the upper end level aligned with the foramen magnum and the lower end level aligned with the lower edge of the first lumbar vertebra in adults. This defined the extent of the ROI, encompassing the entire spinal cord while carefully excluding vertebral and cerebrospinal fluid regions. This approach ensures that significant biomarkers outside the area of spinal cord injury are not overlooked. Figure 1B illustrates our ROI delineation. Consistency between original images and the four filtering techniques ensured uniformity in ROI delineation for each patient.

Each participant's ROIs were independently segmented by two raters, R.W. and Y.W. Intra-rater reliability was assessed by the same radiologist segmenting 30 cases two weeks apart, while inter-rater reliability was evaluated by another radiologist segmenting the same 30 cases. Intraclass correlation coefficients (ICCs) were calculated to assess both intra- and inter-rater reliability, with an ICC > 0.75 indicating strong consistency.

Extraction of Radiomics Features

Using the Pyradiomics module within 3D Slicer, radiomics features were extracted and organized into seven categories for each segmentation in each filtering technique: (1) Shape-based features: 14, (2) First-order features: 18, (3) Gray-level dependence matrix (GLDM) features: 14, (4) Gray-level size zone matrix (GLSZM) features: 16, (5) Neighboring gray-tone difference matrix (NGTDM) features: 5, (6) Gray-level run-length matrix (GLRLM) features: 16, and (7) Gray-level co-occurrence matrix (GLCM) features: 24. Processing with the four filtering techniques yielded a total of 1,070 radiomics features per patient. Detailed feature categories are provided in Supplementary File 3 and4.

Deep learning analysis

Training Deep Learning Models

For training the deep learning models, the sagittal plane image with the maximum ROI for each patient was selected and resampled to 224x224 pixels. Other image data preparation steps followed the same preprocessing process as described above. Three deep learning models were used: ResNet18, ResNet50, and ResNet101, all pretrained on the ImageNet dataset [18]. Five different channels were input into each model: the original image, box, average, Gaussian and median filtering. Each channel underwent an independent yet identical training process with consistent parameter settings. The five separately trained channels ultimately converged to contribute to the classification task. Data were split into training and test sets in a 7:3 ratio. We have carefully set the parameters (initial learning rate: 0.01, batch: 64, 50 epochs). The classification targets were good prognosis (1) and poor prognosis (0).

Model parameters were divided into two parts: 1. the backbone layer (backbone) and 2. the task-specific layer (task-spec). Task-specific parameters were randomly initialized, while backbone parameters utilized the pre-trained model parameters from ImageNet. For the task-specific parameters, the cosine annealing learning rate decay algorithm was applied [19], with details provided in the supplementary file5.

Deep learning feature extraction

The penultimate layer of the model (AveragePooling layer) was used for feature extraction. Our dataset, which includes five categories of images, was independently trained through five identical deep learning channels. Deep learning features were extracted separately for each channel. For ResNet50 and ResNet101 the feature dimensionality was 2048, while for and ResNet18, it was 512. To reduce the risk of overfitting and improve the model's generalization capability, principal component analysis (PCA) was applied to compress the extracted features to 32 dimensions per channel, resulting in 160 deep learning features (32x5) per patient.

In multi-center studies, the ComBat compensation procedure addresses differences arising from imaging protocols and various MRI scanners[18, 19]. This method shows considerable promise in improving the reproducibility of research conducted across multiple centers.

Feature engineering

First, We utilized the ComBat compensation procedure on all extracted features to mitigate multi-center effects[20]. After that, Radiomics features (1,070) were combined with deep learning features (160), resulting in 1,230 fused features per patient. Before feature selection, data were standardized using Z-scores to convert all features to a uniform mean of 0 and a standard deviation of 1.

Initial Screening: Features with a p-value < 0.05 were retained using the Mann-Whitney U test.

Redundancy Reduction: Features were screened for high repeatability using the Spearman rank correlation coefficient, retaining only one feature from any pair with a correlation coefficient greater than 0.9[21]. A greedy recursive deletion strategy was applied to remove the feature with the highest redundancy in each iteration.

Feature Selection: The Minimum Redundancy Maximum Relevance (MRMR) algorithm was used to select the best features. A 10-fold cross-validated Least Absolute Shrinkage and Selection Operator (LASSO) regression identified features with non-zero coefficients for subsequent modeling.

Development and assessment of models

The study employed MLP machine learning methods, implemented using the scikit-learn package (version: 0.18) in Python 3.9. The dataset was randomly split into training (70%) and test (30%) datasets to evaluate model accuracy. Five-fold cross-validation and Grid Search were used to determine the optimal hyperparameters for each model. Evaluation metrics included AUC, Decision Curve Analysis (DCA), Calibration curve, specificity, sensitivity, and accuracy, among others.

Statistical analysis

Categorical variables and normally distributed variables were analyzed using the chi-square test and t-tests, respectively. A p-value < 0.05 indicated statistical significance. Python (version 3.9, http://www.python.org) was utilized for all analyses.

Patient characteristics

A total of 82 patients Cervical SCI were enrolled in this study, who were aged from 9 to 84 years (a mean age of 52 years). As shown in the methods section. The entire dataset was divided into two distinct groups: good prognosis and bad prognosis. Table 1 summarizes the comparison of demographic characteristics of the two groups.

Table 1

Clinical characteristics
Feature_name	Total	Difference in ASI number, no improvement(n = 41)	Difference in ASI number, improvement(n = 41)	pvalue
Age	52.33 ± 12.32	52.83 ± 11.01	51.83 ± 13.63	0.593625
Gender				0.567806
Female	15(18.29)	9(21.95)	6(14.63)
Male	67(81.71)	32(78.05)	35(85.37)
Injury mechanism:				0.082036
Fall on the same level	18(21.95)	13(31.71)	5(12.20)
Fall from a height less than 1 meter	2(2.44)	1(2.44)	1(2.44)
Fall from a height greater than 1 meter	37(45.12)	14(34.15)	23(56.10)
Traffic accident	17(20.73)	7(17.07)	10(24.39)
Others (e.g., sports)	8(9.76)	6(14.63)	2(4.88)
Number of injured segments				0.452569
1	22(26.83)	13(31.71)	9(21.95)
2	36(43.90)	16(39.02)	20(48.78)
3	14(17.07)	6(14.63)	8(19.51)
4	7(8.54)	5(12.20)	2(4.88)
5	1(1.22)	1(2.44)	0
6	1(1.22)	0	1(2.44)
7	1(1.22)	0	1(2.44)
Fracture				0.03609
No	54(65.85)	32(78.05)	22(53.66)
Yes	28(34.15)	9(21.95)	19(46.34)
Preoperative_ASIA_grade				< 0.001
Grade A	13(15.48)	6(14.63)	7(16.28)
Grade B	15(17.86)	4(9.76)	11(25.58)
Grade C	24(28.57)	4(9.76)	20(46.51)
Grade D	26(30.95)	21(51.22)	5(11.63)
Grade E	6(7.14)	6(14.63)	0
Postoperative_ASIA_grade				0.0275
Grade A	7(8.33)	7(17.07)	0
Grade B	7(8.33)	3(7.32)	4(9.30)
Grade C	12(14.29)	8(19.51)	4(9.30)
Grade D	48(57.14)	19(46.34)	29(67.44)
Grade E	10(11.90)	4(9.76)	6(13.95)

Forecasting with the Deep earning and radiomics model

Radiomics features and modeling

A total of 1045 features passed the ICC test. After extensive feature screening, seven features were incorporated into the radiomics model. The model demonstrated an AUC of 0.869 (95% CI 0.7862–0.9525) in the training cohort, with sensitivity (0.875) and specificity (0.727), respectively. In the test cohort, the AUC was 0.7780 (95% CI 0.5285-1), with sensitivity (0.778) and specificity (0.750). Detailed statistical results and curves are presented in Fig. 2 and Table 2.

Table 2

Comparision results of performance different Deep learning models
Model	Cohort	AUC	95% CI	Accuracy	Sen	Spe	PPV	NPV	Pre	Recall	F1
Rad	train	0.8690	0.7862–0.9525	0.8000	0.8750	0.7270	0.7570	0.8570	0.7570	0.8750	0.8120
Rad	test	0.7780	0.5285–1.0000	0.7650	0.7780	0.7500	0.7780	0.7500	0.7780	0.7780	0.7780
ResNet18	train	0.9480	0.9014–0.9945	0.8770	0.8480	0.9060	0.9030	0.8530	0.9030	0.8480	0.8750
ResNet18	test	0.7920	0.5695–1.0000	0.7650	0.8750	0.6670	0.7000	0.8570	0.7000	0.8750	0.7780
Rad+ ResNet18	train	0.8570	0.7703–0.9437	0.7690	0.6250	0.9090	0.8700	0.7140	0.8700	0.6250	0.7270
Rad+ ResNet18	test	0.8470	0.6474–1.0000	0.8240	0.7780	0.8750	0.8750	0.7780	0.8750	0.7780	0.8240
ResNet50	train	0.9750	0.9358–1.0000	0.9550	0.9390	0.9700	0.9690	0.9410	0.9690	0.9390	0.9540
ResNet50	test	0.8750	0.6898–1.0000	0.8750	0.8750	0.8750	0.8750	0.8750	0.8750	0.8750	0.8750
Rad+ ResNet50	train	0.9400	0.8835–0.9971	0.9240	0.9090	0.9390	0.9370	0.9120	0.9370	0.9090	0.9230
Rad+ ResNet50	test	0.9220	0.7919–1.0000	0.8750	0.8750	0.8750	0.8750	0.8750	0.8750	0.8750	0.8750
ResNet101	train	0.9270	0.8682–0.9849	0.8640	0.8790	0.8480	0.8530	0.8750	0.8530	0.8790	0.8660
ResNet101	test	0.8590	0.6692–1.0000	0.8120	0.8750	0.7500	0.7780	0.8570	0.7780	0.8750	0.8240
Rad+ ResNet101	train	0.8820	0.8030–0.9602	0.8150	0.8790	0.7500	0.7840	0.8570	0.7840	0.8790	0.8290

Table 3

MRI Scanners in Three Centers
Item	Fujian Medical University Union Hospital	Ordos Central Hospital	The First People’s Hospital of ChangDe City
MRI Machine	GE Signa	GE Signa	GE Signa
TR Settings (ms)	2500–3000	2100	2000
TE Settings (ms)	100–110	120	80
Slice Thickness (mm)	3	3	3
Slice Spacing (mm)	1	1	1
Matrix Configuration	256 × 512	256 × 512	256 × 512
Field of View (FOV) (mm × mm)	240 × 240	240 × 240	240 × 240
Imaging Plane	Sagittal	Sagittal	Sagittal

Deep learning model

The features used to build ResNet-18, ResNet-50, and ResNet-101 were 10, 9, and 10, respectively. With regard to the predictions generated by the deep learning models, among the three models we utilized, ResNet50 performed the best, followed by ResNet101. In the test set, the AUCs for ResNet50, ResNet101, and ResNet18 were 0.8750, 0.8590, and 0.7920, respectively; their accuracies were 0.8750, 0.8120, and 0.7650; sensitivities were 0.8750, 0.8750, and 0.8750; specificities were 0.8750, 0.7500, and 0.6670. (Fig. 2, Table 2). Supplementary File 6 illustrates the Lasso regression paths and Mean Squared Error (MSE) plots for each model. Supplementary File 7 provides the feature weights utilized in constructing these models.

Combined model

Considering the number of features used to establish each model, the radiomics + ResNet18 and radiomics + ResNet50 models embraced the inclusion of nine features(Figure4), while the radiomics + ResNet101 model was constructed using eight select features.

Among these, the Rad + ResNet50 model also performed the best: In the test set, the AUCs for Rad + ResNet50, Rad + ResNet101, and Rad + ResNet18 were 0.9220, 0.8750, and 0.8470, respectively; their accuracies were 0.8750, 0.8820, and 0.8240; sensitivities were 0.8750, 0.8750, and 0.7780; specificities were 0.8750, 0.8890, and 0.8750(Fig. 2). For more details, please reference Table 2.

DCA provided additional supporting evidence for these conclusions. DCA shows that when the threshold probability surpasses the 20% mark, the model produces more net benefit in the current study. The Decision Curve Analysis (DCA) and calibration plots indicate strong clinical utility. (Fig. 3A-B)

Grad-CAM

Visualizing the final convolutional layer using Grad-CAM is an effective method to enhance model interpretability. As illustrated in Fig. 3C, the darker blue regions highlight crucial areas for prognosis prediction. These regions are primarily focused on the edematous spinal cord, deformed vertebrae, and swollen ligaments.

In this study, we included 82 patients with cervical spinal cord injury (SCI). We discovered that the predictive capability of the feature fusion model was enhanced compared to using radiomics or CNNs individually. Ultimately, we found that the Rad + ResNet-50 model had the highest prognostic predictive value for cervical SCI, with AUCs of 0.940 and 0.922 in the training and test cohorts, respectively.

Contributions of This Study and Comparison with Previous Studies

Functional prognosis after cervical spinal cord injury (SCI), particularly regarding regaining functional independence, is a primary focus in rehabilitation. Predicting SCI recovery can be challenging, but the ASIA grade is a key benchmark for assessing clinical recovery and long-term prognosis [22]. In our study, 82 cervical SCI patients were divided into two groups based on whether their ASIA index improved after one year. Table 1 shows preoperative and postoperative ASIA grades with p < 0.05, indicating clinical significance. Yann Facchinello et al. also highlight that the severity of the neurological deficit at admission, indicated by the ASIA grade, is a critical predictor of recovery [23].

In addition to the ASIA score, MRI images are crucial for prognosis, with metrics shown to be reproducible and predictive of clinical outcomes[24].Hyun-Joon Yoo et al. used machine learning algorithms to create a prediction model for SCI patients based solely on clinical variables, achieving promising results [25]. However, their approach did not include the prognostic value of imaging, particularly MRI.

The BASIC (Brain and Spinal Injury Center) score, developed to predict neurological improvement based on high signal intensity on axial T2-weighted MRI, has proven effective [25] [10]. However, this score focused on a limited set of manually measured MRI features, such as injury length and spinal canal compromise, overlooking many other potential features.

Our study leverages deep learning radiomics to extract and model MRI features. This approach reduces the risk of missing significant features, as evidenced by the combined model (Rad + ResNet50 AUC: 0.9220), which demonstrates improved performance.

We noted that in another study using deep learning, Okimatsu and colleagues developed a CNN model to quantify radiographic characteristics and predict 1-month neurological outcomes in acute SCI patients, achieving 71% accuracy[26]. We extended the follow-up period (1-year) and also improved the model's accuracy to 92.2%. Integrating handcrafted radiomics (HCR) and deep transfer learning (DTL) features for cervical SCI prognosis is novel, and our approach uses standard image data without special training, highlighting its potential.

Training deep learning models on small datasets has limitations, though deep transfer learning can help. André Wirries suggested that using small datasets to train deep learning models for clinical applications is practical[27]. We enhanced model performance through random shifts, rotations, and horizontal flips during training to increase data diversity[28]. Expanding the dataset and controlling feature numbers further improved results. We implemented rigorous feature selection processes in our experiments[29, 30].

ResNet’s structure uses shortcut connections, enabling effective fusion training[31]. To evaluate different models, we tested ResNet-18, ResNet-50, and ResNet-101. Both deep learning and combined models showed that ResNet-50 performed best. Despite having fewer layers and higher Top-1 and Top-5 error rates than ResNet-101[31], ResNet-50’s moderate depth proved advantageous in this research. This indicates that while deeper networks may offer superior learning capabilities, selecting an appropriate network depth for specific research can lead to better model fitting and other benefits, such as reduced computational resources and time. Additionally, ResNet-18 underperformed compared to ResNet-50, suggesting lower-dimensional ResNet networks may lack sufficient fitting capacity.

Deep Learning Model Interpretability

Although deep learning shows promise in disease diagnosis and prognosis, the interpretability of these models can hinder their broader application[32, 33]. Some studies use post-hoc methods or supervised machine learning models to interpret deep learning algorithm outputs[34]. For example, Yixin Wang visualized SVM features using a SHAP-based method[34]. In our study, we generated Grad-CAM images to enhance model interpretability, similar to how Kim, Y. and colleagues used heatmaps to identify regions crucial for prognosis prediction[35].

limitations

There are some limitations to this study that further studied and explored in future work. First, the sample size of cervical SCI was small. Although our findings reflect the predictive ability of deep learning features to a certain extent, more data would be more convincing. Second, this was a restrospective study, more prospective data are needed to verify the effectiveness of the model.

In conclusion, the DLR features were fused with cervical SCI features based on MRI images, which improved the identification ability of a single radiomics prognostic model for cervical SCI. And we developed a new prognostic model (Rad + ResNet-50) which had the highest predictive value for prognosis of cervical SCI with AUCs of 0.940 and 0.922 in the training cohort and test cohort, respectively.

Funding/Support: This research received no external funding

Conflict of Interest: The authors have nothing to disclose.

Research Ethics: Institutional Review Board approval was obtained. The study was conducted according to the guidelines of the Declaration of Helsinki, and was approved by Institutional Review Board (approval no. 2021KY138).

Acknowledgments: We thank the colleagues in our department for their help in our study.

Informed Consent Statement: Written informed consent was not required for this study because this is a retrospective study.

Data availability：The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Alizadeh, A., S.M. Dyck, and S. Karimi-Abdolrezaee, Traumatic Spinal Cord Injury: An Overview of Pathophysiology, Models and Acute Injury Mechanisms. Front Neurol, 2019. 10: p. 282.
Aarabi, B., et al., Intramedullary Lesion Length on Postoperative Magnetic Resonance Imaging is a Strong Predictor of ASIA Impairment Scale Grade Conversion Following Decompressive Surgery in Cervical Spinal Cord Injury. Neurosurgery, 2017. 80(4): p. 610–620.
Wichmann, T.O., et al., Early clinical predictors of functional recovery following traumatic spinal cord injury: a population-based study of 143 patients. Acta Neurochirurgica, 2021. 163(8): p. 2289–2296.
Giger, M.L., Machine Learning in Medical Imaging. J Am Coll Radiol, 2018. 15(3 Pt B): p. 512–520.
Wong, K.K.L., L. Wang, and D. Wang, Recent developments in machine learning for medical imaging applications. Comput Med Imaging Graph, 2017. 57: p. 1–3.
English, M., et al., Machine Learning in Neuro-Oncology, Epilepsy, Alzheimer's Disease, and Schizophrenia. Acta Neurochir Suppl, 2022. 134: p. 349–361.
Dietz, N., et al., Evaluation of Predictive Models for Complications following Spinal Surgery. J Neurol Surg A Cent Eur Neurosurg, 2020. 81(6): p. 535–545.
Inoue, T., et al., XGBoost, a Machine Learning Method, Predicts Neurological Recovery in Patients with Cervical Spinal Cord Injury. Neurotrauma Rep, 2020. 1(1): p. 8–16.
Dietz, N., et al., Machine learning in clinical diagnosis, prognostication, and management of acute traumatic spinal cord injury (SCI): A systematic review. J Clin Orthop Trauma, 2022. 35: p. 102046.
Haefeli, J., et al., Multivariate Analysis of MRI Biomarkers for Predicting Neurologic Impairment in Cervical Spinal Cord Injury. AJNR Am J Neuroradiol, 2017. 38(3): p. 648–655.
Khan, O., et al., Use of Machine Learning and Artificial Intelligence to Drive Personalized Medicine Approaches for Spine Care. World Neurosurg, 2020. 140: p. 512–518.
Azimi, P., et al., A Review on the Use of Artificial Intelligence in Spinal Diseases. Asian Spine J, 2020. 14(4): p. 543–571.
Yin, Y., et al., The Efficacy of Anterior Cervical Corpectomy and Fusion and Posterior Total Laminectomy on Cervical Spinal Cord Injury and Quality of Life. Comput Math Methods Med, 2022. 2022: p. 8216339.
Seng, C., et al., Surgically treated cervical myelopathy: a functional outcome comparison study between multilevel anterior cervical decompression fusion with instrumentation and posterior laminoplasty. Spine J, 2013. 13(7): p. 723–31.
Tarawneh, A.M., et al., Can MRI findings predict the outcome of cervical spinal cord Injury? a systematic review. European Spine Journal, 2020. 29(10): p. 2457–2464.
Zhang, M.Z., et al., Predicting postoperative recovery in cervical spondylotic myelopathy: construction and interpretation of T2(*)-weighted radiomic-based extra trees models. Eur Radiol, 2022. 32(5): p. 3565–3575.
Burns, J.E., J. Yao, and R.M. Summers, Vertebral Body Compression Fractures and Bone Density: Automated Detection and Classification on CT Images. Radiology, 2017. 284(3): p. 788–797.
Orlhac, F., et al., Validation of A Method to Compensate Multicenter Effects Affecting CT Radiomics. Radiology, 2019. 291(1): p. 53–59.
Lucia, F., et al., External validation of a combined PET and MRI radiomics model for prediction of recurrence in cervical cancer patients treated with chemoradiotherapy. Eur J Nucl Med Mol Imaging, 2019. 46(4): p. 864–877.
Bi, S., et al., Multi-parametric MRI-based radiomics signature for preoperative prediction of Ki-67 proliferation status in sinonasal malignancies: a two-centre study. Eur Radiol, 2022.
Wang, W., et al., Development and Validation of a Computed Tomography-Based Radiomics Signature to Predict Response to Neoadjuvant Chemotherapy for Locally Advanced Gastric Cancer. JAMA Netw Open, 2021. 4(8): p. e2121143.
van Middendorp, J.J., et al., Diagnosis and prognosis of traumatic spinal cord injury. Global Spine J, 2011. 1(1): p. 1–8.
Facchinello, Y., et al., Use of Regression Tree Analysis for Predicting the Functional Outcome after Traumatic Spinal Cord Injury. J Neurotrauma, 2021. 38(9): p. 1285–1291.
Andreoli, C., et al., MRI in the acute phase of spinal cord traumatic lesions: Relationship between MRI findings and neurological outcome. Radiol Med, 2005. 110(5–6): p. 636–45.
Yoo, H.J., et al., Prediction of gait recovery using machine learning algorithms in patients with spinal cord injury. Medicine (Baltimore), 2024. 103(23): p. e38286.
Okimatsu, S., et al., Determining the short-term neurological prognosis for acute cervical spinal cord injury using machine learning. J Clin Neurosci, 2022. 96: p. 74–79.
Wirries, A., et al., Artificial intelligence facilitates decision-making in the treatment of lumbar disc herniations. Eur Spine J, 2021. 30(8): p. 2176–2184.
Kuo, B.I., et al., Keratoconus Screening Based on Deep Learning Approach of Corneal Topography. Transl Vis Sci Technol, 2020. 9(2): p. 53.
You, Z., et al., Vocal cord leukoplakia classification using deep learning models in white light and narrow band imaging endoscopy images. Head Neck, 2023. 45(12): p. 3129–3145.
Halligan, S., Y. Menu, and S. Mallett, Why did European Radiology reject my radiomic biomarker paper? How to correctly evaluate imaging biomarkers in a clinical setting. Eur Radiol, 2021. 31(12): p. 9361–9368.
He, K.a.Z., Xiangyu and Ren, Shaoqing and Sun, Jian, Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: p. 770–778.
Ching, T., et al., Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface, 2018. 15(141).
Madabhushi, A. and G. Lee, Image analysis and machine learning in digital pathology: Challenges and opportunities. Med Image Anal, 2016. 33: p. 170–175.
Wang, Y., et al., The radiomic-clinical model using the SHAP method for assessing the treatment response of whole-brain radiotherapy: a multicentric study. Eur Radiol, 2022. 32(12): p. 8737–8747.
Kim, Y., et al., A CT-based Deep Learning Model for Predicting Subsequent Fracture Risk in Patients with Hip Fracture. Radiology, 2024. 310(1): p. e230614.

No competing interests reported.

file.docx

Download PDF

Reviews received at journal
18 Sep, 2024
Reviewers agreed at journal
03 Sep, 2024
Reviewers agreed at journal
30 Aug, 2024
Reviewers invited by journal
28 Aug, 2024
Editor assigned by journal
28 Aug, 2024
Editor invited by journal
14 Aug, 2024
Submission checks completed at journal
12 Aug, 2024
First submitted to journal
02 Aug, 2024

You are reading this latest preprint version

Postoperative One Year Prediction for Patients with Cervical Spinal Cord Injury Based on Deep Learning and Radiomics

Status:

Version 1

Abstract

Background:

Methods:

Results:

Conclusion:

Figures

Introduction

Materials and methods

Patient Selection and Study Design

Surgical Procedures

Clinical Data Collection and ASIA Scale Assessment

Acquisition and Preprocessing of T2-weighted MRI Images

ROI Segmentation and Radiomics Analysis

Segmentation of Regions of Interest (ROI)

Extraction of Radiomics Features

Deep learning analysis

Training Deep Learning Models

Deep learning feature extraction

Feature engineering

Development and assessment of models

Statistical analysis

Results

Patient characteristics

Forecasting with the Deep earning and radiomics model

Radiomics features and modeling

Deep learning model

Combined model

Grad-CAM

Discussion

Contributions of This Study and Comparison with Previous Studies

Deep Learning Model Interpretability

limitations

Conclusions

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1