Spinal Cord Injury Identification and Localization Detection Based on MRI Imaging and Deep Learning Technology

SCI is caused by various reasons leading to structural and functional damage to the spinal cord, resulting in partial or complete loss of motor, sensory, and autonomic nervous functions below the level of injury, primarily due to trauma [1-3]. There are nearly 40 million SCI patients worldwide, with over a million in China alone. The annual incidence rate is 37 per million, increasing by 120,000 each year [4, 5]. SCI not only causes structural and functional damage to the body but also leads to many complications. About 80% of patients suffer from neurogenic bladder and neurogenic bowel dysfunction [6, 7], have a significantly higher risk of venous thrombosis embolism (VTE) [8] and pressure injury (PI) [9] than the general population, and severe respiratory dysfunction has become the leading cause of death in the acute and chronic recovery phases [10]. Therefore, the high disability rate, mortality rate, complication rate, long rehabilitation time, and poor outcome of SCI severely affect the physical and mental health and quality of life of patients [11]. Since there is currently no effective cure for SCI, how to diagnose and differentiate the injury site for SCI patients as soon as possible, quickly clarify and implement treatment plans, reduce the incidence and disability rate of complications, restore bodily functions, and improve the survival quality of patients has always been the focus of SCI research.

MRI of the spine is one of the important means for diagnosing spinal diseases clinically. MRI is highly valuable in observing the shape of the spinal cord, signal intensity, vertebral bone quality, and changes in intervertebral discs [12], and provides clear indications of spinal cord edema and hemorrhage. It not only helps determine the cause and extent of neurological impairment, the possible mechanism of injury, and spinal instability but also provides important auxiliary reference for the diagnosis and differentiation of SCI. In recent years, deep learning technology has achieved good results in the field of medical image recognition, especially in disease detection and identification. For instance, Anantharajan et al. [13] used deep learning and machine learning methods combined with brain MRI images for brain tumor detection, Bousis et al. [14] conducted a survey of deep learning detection of skin cancer research results, organizing the most common deep learning models and datasets used for skin cancer classification, and Lundervold and Lundervold [15] studied the application of deep learning technology from MRI image acquisition to image retrieval, segmentation, and disease prediction. In recent years, more and more clinical departments have combined MRI image datasets with machine learning for the diagnosis and treatment of spinal cord diseases. For example, de Paiva et al. [16] reviewed the most common and noteworthy intramedullary and extramedullary spinal tumors, as well as other tumor mimetics, with a focus on their MRI morphological features. Zhang et al. [17] proposed that machine learning and PPI analysis can help screen and diagnose the basic PRG (pyroptosis related genes) for SCI. Kim et al. [18] developed an advanced predictive model for pressure ulcer occurrence in SCI patients using machine learning (ML) technology. Jazayeri et al. [19] uses machine learning algorithms to develop a region based prediction model for the incidence rate of traumatic spinal cord injury. However, due to the wide variability in the shape, length, and range of lesions of the spinal cord, and issues with image changes in size, resolution, orientation, contrast, and artifacts, their research application in large clinical studies of multi-part images has been limited. Currently, the application of deep learning technology in common spinal diseases on MRI images is relatively scarce. We attempt to use deep learning technology to preliminarily explore the location identification and detection of SCI on MRI images, which could further provide a possibility for its clinical application. With the continuous deepening of clinical research on SCI, there has been significant progress in the mechanisms and medications of SCI, but clinical research still remains at the primitive manual level, with diagnosis of SCI still based on experience and visual estimation, without quantification of the specific size of SCI or specific degree division of the same layer SCI signals. To address the problems in the existing technology, this paper collects MRI imaging data of SCI patients, gathers a dataset, and uses the deep learning technology—Faster R-CNN to recognize, detect, and locate lesions in SCI MRI images, and evaluates deep learning network models, comparing the prediction effects and detection speeds of different network architecture models, verifying the effectiveness of the big data + artificial intelligence model in the identification and detection of SCI diseases, laying a foundation for more extensive research in the field of spinal and spinal cord imaging in the future.

Table 3. Labeling of lesion areas for edematous SCI

Number	Original Image	Spinal Cord Abnormal Signal Label	Annotation
1	31.png	32.png	L4/5 significant disc herniation with free compression of the spinal cord, noticeable edema and high signal above the nucleus pressing on the spinal cord.
2	33.png	34.png	Cervical spondylotic myelopathy, multi-segmental spinal canal stenosis, long-term compression of the spinal cord, MRI indicates local edema with high signal, significant behind the C5 vertebra.
3	35.png	36.png	Cervical conus canal stenosis, C4/5, C5/6 disc herniation, high signal edema behind the 4th cervical vertebra in the spinal cord.

3.3 Image preprocessing

To prevent overfitting in the learning process of visual tasks due to a small amount of data and to further improve the performance of the neural network model, reduce the model's sensitivity to images, and avoid sample imbalance [23], it is necessary to perform data augmentation on image data. First, batch processing is applied to the original MRI images to standardize all images to a size of 600*800, and label extraction is conducted to form an MRI image database. The normalization process uses formula (1) to adjust the image grayscale to the range of 0-255.

$G_1(x, y)=\left(G_0(x, y)-\min \left(G_0\right)\right) \times \frac{255-0}{\max \left(G_0\right)-\min \left(G_0\right)}$           (1)

Among them, G₀(x,y) is the grayscale value of the original MRI image at (x, y) pixels, and G₁(x,y) is the grayscale value of the digitized MRI image at (x, y) pixels after grayscale normalization. All subsequent operations are performed on grayscale normalized images G₁(x,y).

Then, data augmentation, including angle rotation, horizontal flipping, vertical flipping, and random scaling, expands the data to 10 times its original size, with image specifications of 600*800 (length*width*RGB). The ImageDataGenerator module of Python's keras library [24] has several parameters for transforming images in different ways. This study mainly expanded the training set sample data through four methods, as shown in Table 4:

Table 4. Data augmentation methods and parameters

5.1 Choice of feature extraction network

4.png

Figure 4. VGG-16 Network Structure

5.png

Figure 5. ResNet50 network structure and intermediate dimension transformation

6.png

Figure 6. Faster R-CNN network structure

7.png

Figure 7. Recognition and localization process of the Resnet 50 faster R-CNN model

8.png

Figure 8. Workflow of the VGG16 backbone model

The training process is as follows:

(1) Feature Map Extraction: Labeled original images go through the Conv+ReLU+Pooling layers of ResNet50, extracting the Input image's Feature maps. These Feature maps are used for subsequent RPN layers and layers beyond.

(2) RPN: Primarily utilizes a sliding window to connect to the Feature map output by the last convolutional layer of the CNN network, taking the feature map as the input to the output layer. After generating a set of Anchor boxes, a Softmax determines if these Anchors are foreground or background. Signal changes are considered foreground, and the rest as non-foreground, akin to a binary classification problem; during the R-CNN process, the generated proposals undergo IOU evaluation, with those over 0.7 considered foreground, identifying possible lesion locations. Additionally, another branch of the network regresses these Anchor boxes for proposal refinement, aiming for precision.

(3) ROI Pooling: Receives the Feature map from the last layer of the CNN ResNet50 and the refined proposals generated by RPN. After ROI Pooling, a fixed-size proposal feature map is obtained, which is then fed into the subsequent network for target recognition and localization operations using full connections.

(4) Model Tail: Consists of a Softmax classifier and a bounding box regressor (b-box regressor). The fixed-size Feature map formed by the ROI Pooling layer undergoes a full connection operation (compared to R-CNN, SPP Net in classification). Faster R-CNN has two network output layers, integrating the separately operated b-box regression into a unified network in a formal sense and simultaneously constructing a loss function that optimizes both output layers. The network architecture includes four loss functions (rpn_cls_loss, rpn_box_loss, rcnn_cls_loss, rcnn_box_loss), with RPN and Fast R-CNN each having their own classifiers and regressors.

5.4.3 VGG16 faster R-CNN training process

The Faster R-CNN network based on VGG16, compared to the ResNet network, has slight differences in feature extraction by the convolutional layers. The subsequent processes, shown in Figure 8, including the RPN layer, Pooling layer, and full connections for classification and target localization, are essentially the same:

(1) Images are input into a VGG-16 network for feature extraction; resulting in a set of Feature maps, similar to ResNet50, with Feature map=38*50;

(2) Potential lesion locations are determined through the RPN network;

(3) These location information is located on the Feature map through the ROI Pooling layer;

(4) Finally, the region information obtained through the VGG-16 classifier and IOU are evaluated for overlap degree, integrated into a function and backpropagated, completing a training round.

This paper presents a method based on deep learning for SCI analysis, aimed at accurately detecting, identifying, and marking the presence of SCI within MRI images, including their location, size, and prediction probability. This method represents a shift from qualitative to quantitative analysis of SCI, effectively narrowing the diagnostic gap due to varying levels of clinical experience and improving diagnostic accuracy, reducing workload, and overcoming the limitations, randomness, and uncontrollability of SCI diagnosis and treatment. This paper demonstrated that the application of the Faster R-CNN algorithm with neural network models based on VGG-16 and ResNet50 can recognize and detect common diseases such as spinal disc herniation and changes in spinal cord signals. Using standardized datasets, selecting appropriate models, and evaluation criteria can lead to satisfactory experimental results. However, the network architecture model based on ResNet50 outperforms the one based on VGG-16 in terms of predictive effect and detection speed, indirectly proving that deeper network structures help improve prediction effects. It also confirms the theoretical effectiveness of the big data + artificial intelligence model.

The innovative content of this article is as follows: (1) Using the current hot topic of deep learning technology in artificial intelligence - Faster R-CNN, to explore whether it can recognize and detect common cervical spine diseases in magnetic resonance imaging, in order to verify the feasibility and effectiveness of deep learning technology for diseases in this field. (2) This paper adopts the method of object recognition detection in deep learning to locate and predict common cervical spine diseases, and compares experiments with different network structures. It is found that the deeper the neural network, the better the training effect. (3) The successful recognition and detection of common diseases in the cervical spine using deep learning technology has laid the theoretical foundation for clinical application of cervical spine disease MRI+deep learning mode. Of course, there are still certain challenges from theoretical foundations to model integration and practical clinical workflows, such as accuracy, transparency, and interpretability, as well as concerns about data privacy and security. In terms of accuracy, the training data of deep learning models for spinal cord injury detection is relatively scarce due to the involvement of privacy, and the accuracy needs to be improved. During the training process of the large model, it is necessary to strengthen data validation, increase uncertainty indicators, optimize medical accuracy, and self improve through error correction algorithms and other methods. In terms of transparency and interpretability, it is not yet clear how the model generates answers (black box questions) from input queries and data structures, nor is it clear which parts of the training dataset are used in the result response. To solve this problem, it is necessary to reference the data section that contributes to the answer in the model output, and conduct in-depth research and development of interpretable AI. The application of large models in the medical field faces multiple technical challenges and limitations, such as insufficient knowledge, limited interpretability and accuracy, coherence intersections, and model hallucinations. Overcoming these issues requires interdisciplinary collaboration, strengthening data management and protection, researching interpretable AI methods, and continuously improving the performance and security of models to ensure their reliability and effectiveness in medical practice.

Certainly, this paper has some limitations, such as in data collection, where there may be issues like insufficient structural quality of data, small data volume, incomplete indication data, and difficulty in sharing. The completeness and accuracy of data can affect the performance of the neural network model; the deep learning model is still not perfect, and there is room for improvement in detection accuracy. Future research directions should continue to investigate the causes of misdiagnosis and missed diagnosis, consider designing a correction module, and use correction algorithms to correct the automatic detection results of the trained spinal cord analysis neural network. After image detection is complete, a comparison of the signal differences recognized by T2WI and T2 fat suppression should be conducted, followed by a comparison at the same level and dimension. After the comparison, match analysis weight to correct instances where some interferences are mistakenly identified as SCI.