A review of the application of novel intervertebral disc diagnostic technologies integrated with artificial intelligence in medical imaging

2.1 Applications and limitations of conventional X-ray and CT in disc diagnosis

Conventional X-ray imaging is widely employed as an initial diagnostic modality for intervertebral disc diseases, its primary value lies in assessing spinal alignment (e.g., scoliosis, spondylolisthesis) and osseous changes, It can effectively reveal imaging manifestations such as intervertebral space narrowing and spondylolisthesis, which may indicate intervertebral disc degeneration. ¹⁰ Owing to its operational convenience and low cost, X-ray is frequently used as a preliminary screening tool in outpatient clinics.

In traditional diagnostics, physicians must rely on their own expertise to correctly identify spinal segments and radiographic projections, manually annotate key anatomical landmarks, and perform a series of geometric measurements to obtain quantitative assessment data. ¹¹ They also need to search for and recognize various imaging features of degenerative changes based on specific positional views. Ultimately, a clinical diagnosis is made by integrating all measured data and observed characteristics. Nevertheless, it exhibits considerable limitations: its poor soft-tissue resolution prevents direct visualization of disc herniation or neural compression, resulting in low detection rates for early degenerative changes and subtle lesions (e.g., non-calcified herniations), with reported sensitivity as low as 51%–61% in diagnosing lumbar pathologies. ¹² Meanwhile, the manual diagnostic process is time-consuming and suffers from relatively low accuracy. One study indicated that traditional manual diagnosis based on X-rays achieved an accuracy of only 68.3% in diagnosing cervical spondylosis. ⁹ Additionally, the use of ionizing radiation also limits its suitability for repeated examinations.

CT provides high spatial resolution of bony anatomy, allowing detailed evaluation of vertebral osteophytes, pedicle morphology, and spinal canal stenosis. Coupled with 3D reconstruction techniques, CT enables precise visualization of vertebral structures and pedicles, offering critical guidance for surgical interventions, including the planning of percutaneous endoscopic approaches. ¹³ Traditional CT diagnosis involves two primary scanning protocols: direct axial scanning, and spinal helical scanning followed by manual post-processing reconstruction. The former requires stepwise adjustment of the gantry tilt to align the scanning plane with the target intervertebral disc. However, it sometimes fails to adequately visualize the disc and its adjacent structures due to conditions such as scoliosis, carrying a risk of misdiagnosis. The latter protocol, while eliminating the need for disc-by-disc angle adjustment, requires manual, slice-by-slice reconstruction of standardized axial disc images, which are then subjectively analyzed by radiologists. The entire workflow is cumbersome, time-consuming, and labor-intensive. Moreover, even on these reconstructed images, the accuracy of manual detection for the presence of disc herniation (Figure 4) and its subsequent classification reaches only 77.16%. ⁴ Meanwhile, its soft-tissue contrast remains insufficient to discriminate between the nucleus pulposus and annulus fibrosus within the disc. Furthermore, CT entails substantially higher radiation exposure compared to X-ray, raising concerns about cumulative dose risks with repeated imaging. ¹⁰ OPEN IN VIEWER

In summary, while both X-ray and CT provide valuable information regarding osseous abnormalities, neither modality adequately captures the full spectrum of pathophysiological changes in disc degeneration. Moreover, the manual diagnostic process is labor-intensive, time-consuming, and potentially prone to significant variability. This highlights the need for more advanced assistive technologies to help overcome these considerable limitations.

2.2 Applications and limitations of MRI in disc diagnosis

MRI is widely regarded as the “gold standard” for diagnosing intervertebral disc diseases, owing to its superior soft tissue resolution. Through multi-sequence protocols—such as T1-weighted and T2-weighted imaging—MRI clearly delineates disc anatomy (including the nucleus pulposus and annulus fibrosus), hydration status, and adjacent soft tissue structures (e.g., nerve roots and thecal sac).^10,14 It effectively captures microstructural changes associated with disc degeneration, including alterations in signal intensity of the nucleus pulposus, integrity of the annulus fibrosus, and the presence of nerve root compression. ¹⁵ Furthermore, MRI can accurately depict the compression of soft tissues in cases of disc herniation, thereby aiding in its diagnosis (Figure 5). The diagnostic workflow for radiologists when interpreting MRI typically involves: detecting and localizing the affected intervertebral disc; differentiating and grading disc degeneration (using systems such as the Pfirrmann grading scale ¹⁶ ); classifying the type of herniation (e.g., protrusion, extrusion, sequestration) and its location (e.g., central, foraminal); and assessing the severity of secondary complications, such as central canal stenosis and nerve root compression. Additionally, as a non-ionizing modality, MRI is also well-suited for longitudinal monitoring and safe for use in sensitive populations, including pregnant women.OPEN IN VIEWER

Despite these advantages, MRI has several limitations. Although exceptional in soft tissue contrast, it underperforms compared to CT in visualizing calcified tissues, often necessitating complementary CT imaging for cases involving ossified or calcified disc herniations. ¹⁷ Beyond this, conventional MRI biomarkers such as the high-intensity zone (HIZ) exhibit limited sensitivity—only 62.8% in detecting annular fissures—which may contribute to underdiagnosis of early degenerative changes. ¹⁸ Other constraints include relatively long acquisition times, contraindications in patients with certain metallic implants, and a dependence on radiologist expertise that introduces potential subjectivity and variability in the interpretation of complex or atypical cases.

More importantly, the traditional manual interpretation process itself faces significant bottlenecks in both efficiency and accuracy. When confronted with vast quantities of images, the tasks of manual measurement, localization, and classification by physicians are not only time-consuming and labor-intensive but also lack standardized and quantitative evaluation criteria. There may be diagnostic inconsistencies among different physicians, with studies indicating that such variability can reach up to 15%, ¹⁹ primarily stemming from subjective differences in assessing imaging features. Secondly, the accuracy of manual diagnosis is highly dependent on the type of lesion and the imaging modality, and it is generally lower than that of AI-assisted systems. Research has shown that for identifying calcified lumbar disc herniation on lateral lumbar magnetic resonance imaging (MRI), the accuracy of manual recognition was only 70.87%, which is substantially lower than the 91.67% accuracy achieved by the AI model. ²⁰

To summarize, while MRI offers unparalleled capabilities in soft tissue evaluation and avoids ionizing radiation, its inherent limitations, coupled with the inefficiencies, inaccuracies, and subjectivity of human evaluation, highlight the need for advanced complementary approaches—including AI-enhanced analysis—to improve diagnostic precision, efficiency, and consistency in the evaluation of intervertebral disc diseases.

3.1 Technical foundation

AI refers to a branch of computer science dedicated to developing systems capable of performing tasks that traditionally require human intelligence. ²¹ Broadly speaking, AI methodologies can be classified into three categories ²² : symbolic approaches (e.g., rule-based expert systems), Bayesian methods (relying on probabilistic reasoning), and connectionist models (based on neural networks). In terms of technical subfields, it encompasses domains such as Natural Language Processing (which focuses on enabling computers to comprehend and generate human language), Computer Vision (concerned with interpreting images and videos, including tasks like object detection and facial recognition), DL, ML and so on. Among this, ML, a central subset of AI, entails the design of algorithms that improve automatically through experience. ²³ These systems learn mappings between inputs and outputs directly from data without explicit instruction, enabling applications in prediction, recognition, and decision-making.

Based on the learning paradigm, ML can be divided into four categories ²⁴ :Supervised learning utilizes labeled datasets to learn input-output relationships, supporting tasks such as classification and regression (e.g., artificial neural networks); Unsupervised learning identifies hidden patterns or groupings within unlabeled data (e.g., clustering), which is applicable to exploratory data analysis; Semi-supervised learning combines both labeled and unlabeled samples, often blending techniques from the above categories; Reinforcement learning involves an agent learning optimal behaviors through environmental feedback and reward maximization.

In the field of medical imaging, ML enables quantitative analysis of image features, such as texture and morphology, with radiomics playing a significant role. Radiomics transforms conventional image data into minable high-dimensional data by high-throughput extraction of a large number of quantitative features from medical images, thereby revealing subtle characteristics that are difficult to recognize with the human eye and providing more precise information for clinical practice. Studies have demonstrated that ML-based radiomics models applied in conventional imaging can effectively differentiate between primary and secondary lesions, outperforming traditional radiological evaluation. ²⁵

DL, a rapidly advancing subfield of ML, employs deep neural networks with multiple hidden layers to model complex hierarchical representations. ²⁶ Commonly used architectures include CNNs, ²⁷ Recurrent Neural Networks (RNN), Generative Adversarial Network (GAN) and more recently, Transformers based on self-attention mechanisms. CNNs—among the most widely used architectures—typically consist of convolutional layers (for local feature extraction), pooling layers (for dimensionality reduction), nonlinear activation functions (enable the network to learn and model complex nonlinear relationships), and fully connected layers (the extracted features were aggregated and subsequently utilized to perform the final classification or regression tasks). ²⁸ Compared to traditional ML, DL excels at processing large-scale unstructured data such as images and text, autonomously learning discrimin features without manual engineering. In medical imaging applications, DL has demonstrated remarkable performance in analyzing complex image datasets, capturing subtle diagnostic cues beyond human perception.^29,30 However, these models generally demand greater computational resources and are often considered “black boxes” due to their limited interpretability. ³¹

3.2 Current state of AI technology in spinal imaging

The integration of AI into spinal imaging has catalyzed a profound paradigm shift from traditional qualitative morphological observation toward a highly intelligent and quantitative diagnostic framework. By leveraging ML and DL algorithms, AI enables automated and precise image analysis, providing clinicians with efficient diagnostic tools that significantly enhance both the accuracy and efficiency of spinal evaluations. ³² At the frontend of the clinical workflow, the frontiers of AI research have extended into the optimization of image acquisition and reconstruction protocols. DL-based reconstruction algorithms, such as GAN-based compressed sensing techniques, can substantially accelerate MRI acquisition speeds or reduce radiation exposure in CT imaging without compromising diagnostic fidelity, facilitating high-throughput screening and early intervention.^32,33

At the foundational level, AI technologies facilitate the rapid automated localization and precise segmentation (Figure 6) of critical anatomical units, including the vertebrae, intervertebral discs, and neural structures. By establishing reliable anatomical benchmarks, these automated approaches effectively circumvent the inherent limitations of manual annotation. ³⁴ OPEN IN VIEWER

Beyond structural parsing, AI demonstrates significant potential in computer-aided detection and pathological recognition. In the context of spinal trauma, algorithms assist in identifying vertebral fractures on radiographs or CT scans while providing objective quantitative metrics to enhance the reproducibility of clinical assessments.^35,36 In spinal oncology, specialized models are designed to improve the sensitivity and specificity of metastatic lesion detection on MRI, facilitating the discrimination of malignant pathologies within complex backgrounds to minimize diagnostic omissions and false positives.^37,38 Furthermore, for pervasive degenerative conditions, AI is increasingly utilized to automate the identification and standardized grading of disc degeneration, herniation, and spinal stenosis, bridging the gap between imaging findings and standardized clinical reporting.

4.1 Application of AI combined with X-ray and CT in disc diagnosis

4.1.1 X-ray-based inference of disc pathology via “indirect signs”

Conventional X-ray infers disc pathology primarily through secondary osseous changes, however, this process is inherently constrained by the modality’s limited sensitivity and the subjectivity of manual assessment. AI-driven computational approaches now play an essential role in extracting objective, quantifiable biomarkers from these indirect features.

In addressing disc space narrowing—a key surrogate marker of degeneration—a DL framework integrating High-Resolution Network (HRNet) and Deformable Convolution (DAC) achieved end-to-automated lumbar intervertebral disc height measurement, reporting an intraclass correlation coefficient of 0.93–0.98 against radiological standards. ¹¹ This illustrates the ability of structurally-aware architectures to capture fine-grained anatomical details under projective distortion. Clinically, this automated measurement provides an objective longitudinal tool for monitoring degenerative progression, addressing the subjectivity inherent in manual interpretation. Further exploiting transfer learning, a VGG-16 model optimized for lateral (LAT) cervical X-rays attained 95.8% sensitivity in identifying disc space narrowing and osteophyte formation, highlighting how domain-adapted convolutional networks can excel even with limited disc-level signal. ³

For curvature-based assessment, MVC-Net ⁴¹ (Figure 7) introduced a multi-view correlation mechanism that combines information from anteroposterior (AP) and LAT projections to mitigate vertebral occlusion. This mechanism mimics a radiologist’s spatial reasoning by cross-referencing orthogonal views, enabling the automated calculation of the Cobb angle and sagittal alignment parameters. These measurements are clinically essential for diagnosing spinal deformities such as scoliosis or abnormal kyphosis, which serve as critical indirect indicators of underlying disc stress. Complementing this, a Mask R-CNN-based pipeline achieved real-time vertebral instance segmentation using region-based feature alignment, providing pixel-level anatomical grounding for fully automated Cobb angle measurement while significantly reducing inter-observer variability. ⁴² The SpineTK system ⁴³ further introduced a hardware-invariant calibration technique that ensures measurement consistency across different imaging devices.OPEN IN VIEWER

Figure 7.

Schematic of the Multi-view Correlation Network (MVC-Net) for automated spinal curvature assessment. The model takes anteroposterior (AP) and lateral (LAT) radiographs as input. The core X-module explicitly learns a joint feature representation from both views to address vertebral occlusion in the LAT view. These enriched features are then fed into two parallel output branches: one for spinal landmark estimation (vertebral corner detection in both views) and another for direct Cobb angle estimation. The predicted landmarks are used to calculate key coronal (e.g., Cobb angle) and sagittal (e.g., thoracic kyphosis, lumbar lordosis) parameters, enabling comprehensive and automated AIS diagnosis and severity assessment.

At a higher integrative level, a ResNet-34 architecture fused multi-angle radiographic features—including disc space metrics, alignment parameters, and osteophyte profiles—into a unified diagnostic model, achieving 89.7% accuracy in detecting cervical spondylosis against a manual accuracy baseline of 68.3%. ⁹ This underscores the capacity of DL to synthesize heterogeneous and weak radiographic signals into clinically actionable predictions (Table 1).

OPEN IN VIEWER

Table 1.

Key AI models for intervertebral disc localization and Multi -Angle disc degeneration assessment.

Authors	PublicationYear	Model(s)	Function	Innovation	Main ImagingModality Used
Wu H.et al. 41	2018	MVC-Net	Automated comprehensive assessment of Adolescent Idiopathic Scoliosis	Developed an innovative multi-view feature fusion architecture (termed X-Module), which employs weighted spatial summation and channel-wise concatenation to explicitly model the anatomical correlation between AP and LAT radiographic views. This module effectively learns integrated feature representations that significantly enhance localization accuracy and discriminative capability for end-vertebrae regions.	X-ray
Pan Y.et al. 42	2019	Mask R-CNN	Detection and segmentation of the spine and vertebrae	Employing a dual-model framework based on Mask R-CNN, we implemented separate networks for spine-level and vertebra-level detection and segmentation. This design explicitly addresses two key limitations of single-model approaches: first, it mitigates training inefficiency caused by the significant imbalance in annotation volumes between spine and vertebra labels; second, it prevents erroneous filtering of vertebral instances by the Non-Maximum Suppression algorithm when they are detected within spine regions, substantially improving segmentation accuracy and robustness.	X-ray
Das P.et al. 39	2021	RIMNet	Automated simultaneous recognition and segmentation of Intervertebral Discs in multi-modal MR images	Proposed a region-to-image matching strategy utilizing a dual-stream input architecture that concurrently processes both the original spinal MRI and its corresponding annotation map. This framework incorporates dedicated feature fusion and cross-modality matching mechanisms, significantly enhancing recognition accuracy and enabling simultaneous localization and segmentation of intervertebral discs.	Multi-modal
Suri A.et al. 5	2021	SpineTK	Automated segmentation and analysis of vertebrae and intervertebral discs in multi-modal spinal images	The first comprehensively validated DL system enables high-precision vertebral segmentation across MR, CT, and X-ray imaging modalities. Its cascaded architecture—comprising sequential feature extraction, target localization, and pixel-level segmentation stages—progressively refines the region of interest, significantly improving segmentation accuracy and computational efficiency.	Multi-modal
Maras Y.et al. 3	2022	VGG-16	Automated identification and classification of lesions in cervical spine X-ray films	Compared to other contemporary architectures, VGG-16 replaces large-kernel convolutional filters with stacks of smaller 3×3 kernels. This design significantly increases network depth while reducing parameter count. The sequential combination of these small convolutional kernels achieves receptive fields equivalent to those of larger kernels, while substantially enhancing the model’s nonlinear expressive capacity—a critical advantage for capturing subtle pathological features in cervical spine X-ray imaging.	X-ray
Suri A.et al. 43	2023	SpineTK	Automated measurement of Cobb angle in scoliosis patients	Overcoming a key limitation in existing research, this model remains effective even for patients with surgical hardware such as fixation screws, rods, or pacemakers—populations typically excluded from prior studies. It achieves stable Cobb angle measurements in the presence of such obstructions, with error rates showing no statistically significant difference compared to hardware-free cases (P=0.80).	X-ray
Yao H.et al. 11	2024	DL model based on HRNet combined with DAC method	Automated measurement of anatomical parameters of vertebrae and intervertebral discs on LAT lumbar radiographs	Developed a dual-stage global optimization architecture that enhances HRNet by incorporating global hierarchical structural constraints. This integration explicitly models the positional relationships between vertebrae, enabling the model to prioritize anatomical spatial context and significantly reduce vertebral separation and misdetection errors.	X-ray
Xie Y.et al. 9	2024	ResNet-34	Automated diagnosis and multi-view image classification of Cervical Spondylosis	This approach overcomes a key limitation in prior research by integrating multiple radiographic views—including AP, LAT, and bilateral oblique projections—into a unified learning framework. The model captures disease-specific features across complementary angles, significantly improving diagnostic comprehensiveness and robustness.	X-ray

In summary, recent AI methodologies significantly advance X-ray-based disc diagnosis by combining precise anatomical segmentation with multi-feature integration and cross-view reasoning. These technical developments not only enhance objective quantification of indirect signs but also establish a new paradigm for automated spinal pathology assessment in routine radiography.

4.1.2 Application of AI combined with CT in disc diagnosis

Current research on AI-integrated CT for disc diagnosis remains relatively limited, yet emerging studies indicate transformative potential across multiple dimensions.

A CNN-based system enabled automatic segmentation and Modic change (MC) classification in lumbar disc CT images, demonstrating high segmentation accuracy and diagnostic performance while enhancing radiologists’ interpretive efficacy. ⁴ In the surgical domain, AI-generated 3D fusion images combining MRI and CT data successfully simulated a full endoscopic transforaminal discectomy (FED-TF) approach. This provides intuitive visual guidance for assessing the safe zone of Kambin’s triangle and identifying bony obstructions, which is critical for minimizing nerve root injury and optimizing portal placement during complex surgical planning. ⁴⁴ Further advancing classification performance, a Vision Transformer (ViT) framework outperformed conventional CNNs in both vertebral localization and disc abnormality classification on CT, while providing improved model interpretability through Grad-CAM-generated attention maps. ⁴⁵ (Figure 8) Additionally, an active contour-based AI segmentation system quantitatively evaluated treatment outcomes in lumbar disc herniation (LDH) by measuring disc height reduction and vertebral slippage, enabling precise comparison of therapeutic efficacy ⁴⁶ (Table 2).OPEN IN VIEWER

Figure 8.

A Vision Transformer (ViT) Framework for Lumbar Disc Herniation Diagnosis and Interpretability Analysis in CT Imaging. This figure demonstrates the full pipeline of an automated lumbar disc herniation diagnostic model based on a ViT. The stages are: (1) Preprocessing: Original lumbar CT images undergo normalization and resizing. (2) Patching and Embedding: Images are divided into 32×32pixel patches, which are then linearly projected and combined with positional encodings to form a sequence of embedding vectors. (3) Feature Extraction: The embedded sequence is processed by a Transformer encoder composed of 12 multi-head self-attention blocks. This architecture globally models spatial dependencies among multiple vertebrae. (4) Dual-Task Output: The model features two parallel output branches: a localization branch identifies the specific intervertebral disc level, and a qualitative classification branch determines its status as normal, bulging, or herniated. (5) Interpretability Analysis: Gradient-weighted Class Activation Mapping (Grad-CAM) generates heatmaps to visualize the key image regions (highlighted) that the model relies on for decision-making. The focused areas are consistent with anatomical landmarks used in clinical diagnosis, enhancing the model’s trustworthiness and transparency.

OPEN IN VIEWER

Table 2.

Key AI models for pathological classification and diagnosis of intervertebral disc diseases.

Authors	PublicationYear	Model(s)	Function	Innovation	Main ImagingModality Used
Fan X.et al. 46	2022	AI-enhanced CT image analysis system based on active contour segmentation algorithm​	CT image analysis for LDH	Developed a novel multi-algorithm fusion framework through deep integration of Local Binary Fitting, PreWork, and active contour segmentation, achieving exceptional balance between computational efficiency and segmentation accuracy—overcoming a key limitation of conventional single-algorithm approaches in medical image analysis.	CT
Yamada K.et al. 44	2023	DL-based 3D image fusion model	Preoperative feasibility assessment for FED-TF	Multi-angle path assessment improves surgical safety and expands the methodology for assessing indications for minimally invasive surgery.	CT
Wan L.et al. 4	2025	CNN-based combined model	Automated processing and diagnosis of lumbar disc CT images	A U-Net-based architecture enables automatic reconstruction of high-resolution axial lumbar disc images from spiral CT data, supporting automated diagnosis of disc herniation and MC classification with radiologist-level accuracy.	CT
Chu Q.et al. 45	2025	ViT-based image analysis system	Vertebra localization, Disc abnormality classification	A CT-ViT-based framework was first proposed for LDH diagnosis, utilizing self-attention mechanisms to model complex spatial relationships and global dependencies within imaging data. Grad-CAM visualizations across transformer layers provide interpretable saliency maps for pathological disc features.	CT

In conclusion, the integration of AI with CT imaging not only elevates diagnostic precision and operational efficiency but also actively supports surgical planning and outcome assessment. Despite the currently modest volume of studies, the field exhibits considerable potential for further technical innovation and clinical translation.

4.2 Application of AI combined with MRI in disc diagnosis

4.2.2 Diagnosis of disc herniation

Disc herniation represents one of the most prevalent conditions in spinal practice. The integration of AI—particularly through DL and advanced image recognition technologies—into MRI-based evaluation of LDH has enabled automated, high-fidelity analysis of imaging data, substantially improving both the efficiency and accuracy of diagnostic processes.

The diagnostic capability of AI in LDH has been validated across multiple studies. A DL model utilizing the PP-YOLOv2 algorithm achieved a mean average precision (mAP) of 90.08% in distinguishing among normal discs, LDH, and spondylolisthesis in lumbar MRI, with LDH-specific precision reaching 91.74%. ⁴⁸ The diagnostic value lies in its ability to precisely delineate the boundaries of disc displacement and the degree of spinal canal compromise, helping clinicians differentiate between stable protrusions and high-risk extrusions that may require urgent intervention. For specific subtypes such as calcified lumbar disc herniation (CLDH), a ResNet34-based model (Figure 10) attained accuracies of 91.67% and 88.76% in the internal test set and an external validation cohort, respectively, demonstrating robust generalizability. ²⁰ Further advancing model architecture, the GE-YOLOv8 framework (Figure 11) incorporated a Gradient Search (GS) module and Efficient Channel Attention (ECA) mechanism, achieving superior accuracy and operational efficiency compared to both conventional models and manual diagnosis. ⁴⁹ Even beyond human medicine, a two-stage AI model with dedicated spine localization achieved a mAP of 75.32% in detecting disc herniation in veterinary MRI, illustrating the cross-species applicability of DL approaches. ⁵⁰ A systematic review corroborates that CNN and YOLO-based models frequently exceed 85% accuracy in LDH diagnosis, highlighting considerable potential to standardize interpretations and reduce inter-reader variability, despite ongoing challenges related to limited dataset size and insufficient external validation. ¹⁹ OPEN IN VIEWER

Figure 10.

Schematic architecture of a deep Residual Neural Network (ResNet). The diagram illustrates the key innovation of ResNet: the residual block (detailed in the top inset). Each residual block utilizes shortcut connections that bypass one or more convolutional layers (e.g., 3×3 conv), allowing the input to be added directly to the output of the block. This identity mapping mitigates the vanishing gradient problem in very deep networks, enabling stable training and superior performance. The overall architecture begins with an initial convolution and pooling layer, followed by a series of stacked residual blocks with increasing feature dimensions (e.g., 64, 128, 256), and concludes with global average pooling and a fully connected output layer.

OPEN IN VIEWER

Figure 11.

Architectural comparison between the baseline YOLOv8 model (left) and the proposed GE-YOLOv8 model (right) for LDH detection. Key improvements in the GE-YOLOv8 model include: (1) Replacement of the C2f module with a GS module to enhance multi-scale feature extraction while reducing computational complexity; (2) Integration of an ECA module within the head network to optimize feature channel weights and improve sensitivity to small lesions; (3) Retention of the anchor-free double-branch head structure for efficient bounding box regression and classification. These modifications collectively contribute to the model’s superior performance in accuracy and efficiency.

In summary, AI technology—aided by diverse and optimized model architectures—delivers highly efficient and precise MRI diagnosis of LDH, offering substantial clinical utility. Future efforts should prioritize large-scale, annotated multi-center datasets and rigorous external validation to facilitate translation into routine practice.

4.2.3 Classification and grading of degenerative disc diseases

Lumbar disc degeneration represents a leading etiology of chronic low back pain. Although MRI serves as the gold standard for diagnosis, conventional manual interpretation is often limited by subjectivity and inefficiency. Recent advances in AI, particularly DL models, have demonstrated substantial potential in automating the classification and grading of degenerative disc diseases on MRI, thereby offering precise and standardized solutions for clinical evaluation.

CNNs have been employed to detect multiple degenerative changes—such as disc herniation and bulging—within lumbar MRI studies, while simultaneously performing accurate disc localization and labeling. This approach validates the feasibility of using a unified DL architecture for multi-pathology detection. ⁵¹ Multi-task learning frameworks further extend the functionality of AI, enabling concurrent diagnosis of several related conditions. For instance, one multi-task model achieved accuracies exceeding 80% for grading LDH, lumbar central canal stenosis (LCCS), and lumbar nerve root compression (LNRC) on an internal test set, and maintained performance between 74.1% and 79.6% on an external validation set, showing strong agreement with clinical standards. ⁵² Cross-sequence generalization has also been realized through innovative model designs. The YOLOv7-WRN-SVM model successfully predicted T1ρ-based disc degeneration stages using conventional T1-weighted MR images, attaining an accuracy of 84.0%. This method provides a promising alternative for T1ρ-MR applications in intervertebral disc degeneration (IDD) without requiring specialized sequences. ⁵³

In fine-grained grading tasks, YOLOv8-based models excelled in classifying both severity (4 grades) and spatial distribution (8 categories) of lesions, achieving kappa coefficients of 0.88 and 0.77, respectively, which reflects a high level of detail capture and diagnostic consistency. ⁵⁴ Ensemble methods such as WDRIV-Net—which integrates DenseNet169 and ResNet101 through weighted fusion—attained a classification accuracy of 96.25% for single and combined degeneration types, with Area Under the Curve (AUC) improvements of ≥2% compared to individual models, significantly outperforming conventional approaches. ⁵⁵

Quantitative analysis has been enhanced through models like BianqueNet (Figure 12), which combines signal intensity and geometric parameters (e.g., disc height index) to objectively assess degeneration. The extracted parameters showed strong correlation with Pfirrmann grades, establishing an imaging biomarker framework for degeneration severity. ⁵⁶ CNNs also improved grading reliability across different systems, achieving a kappa (It is a statistical metric used to assess classification agreement, particularly for measuring the degree of consistency between two or more observers when classifying the same set of subjects. It quantifies the agreement beyond that expected by chance alone, with a value greater than 0.60 generally indicating substantial agreement.) of 0.68 in Pfirrmann grading—substantially higher than the average human performance (0.38)—and demonstrated similar benefits in Fujiwara grading for facet joint degeneration. ⁵⁷ OPEN IN VIEWER

Figure 12.

Schematic of an advanced DL architecture for fine-grained intervertebral disc analysis. The model employs an encoder-decoder structure. The Encoder Path extracts multi-scale features through initial convolution/pooling and residual blocks. The ST-SC Module (likely Spatial Transformer -Self-Calibration or similar) utilizes window-based multi-head self-attention (W-MSA) and multilayer perceptrons (MLP) to model long-range spatial dependencies. The DFE Module captures multi-context information using Spatial Pyramid Pooling (SPP) and Atrous Spatial Pyramid Pooling (ASPP). The MFF Module integrates high- and low-level features before transposed convolution and 2x upsampling reconstruct the high-resolution output for precise segmentation or grading.

Notably, AI applications have extended beyond human medicine. One model achieved 85.2% sensitivity in grading canine lumbar disc degeneration, slightly below human-oriented models (89.4%) yet still providing a standardized tool for translational research in animal models ⁵⁸ (Table 3).

OPEN IN VIEWER

Table 3.

Key AI models for quantitative assessment and grading of intervertebral disc degeneration.

Authors	PublicationYear	Model(s)	Function	Innovation	Main ImagingModality Used
Lehnen N C.et al. 51	2021	CoLumbo	Identify lumbar spine anatomical structures; Detect multiple lumbar degenerative pathologies	This approach overcomes the limitations of conventional single-pathology detection systems by enabling simultaneous automated identification and diagnosis of multiple lumbar degenerative pathologies.	MRI
Su Z H.et al. 52	2022	ResNet-50 based multi-task classification model	Automated grading of LDH, LCCS, and LNRC	This work presents the first comprehensive framework for multi-task classification and grading of LDH, LCCS, and LNRC from sagittal MR images. A shared-backbone architecture captures implicit pathological relationships among these conditions, enabling synergistic joint learning and improved diagnostic consistency.	MRI
Zheng H D.et al. 56	2022	BianqueNet	Achieves high-precision segmentation of MC-related areas in lumbar spine T2-weighted MRI and performs quantitative analysis based on the segmentation results.	This model incorporates three novel modules for enhanced segmentation: a Dense Feature Extraction (DFE) module that captures multi-scale semantic information to improve contextual reasoning; a Spatial Transformer and Self-Calibration (ST-SC) module that preserves structural details via residual connections and layer normalization, substantially boosting accuracy along disc and vertebral boundaries; and a Multi-scale Feature Fusion (MFF) module that effectively integrates hierarchical features. The synergistic operation of these components yields segmentation performance markedly superior to the DeepLabv3+ baseline.	MRI
Waldenberg C.et al. 18	2022	AI-based disc fissure detection system​	Detection of annular fissures	This approach integrates texture feature analysis with an attention-driven occlusion mechanism to detect and localize disc fissures imperceptible on conventional MRI. An innovative attention submodel systematically occludes image regions to generate interpretable heatmaps of fissure locations, significantly improving detection sensitivity and providing explicit spatial explanations for model predictions	MRI
Lagerstrand K.et al. 47	2022	ML model-based disc fissure detection system	Classification of outer annular fissures and analysis of correlation with pain	This approach employs a Random Forest-based pipeline for end-to-end automated analysis of lumbar MRI, facilitating seamless processing from feature extraction to disc fissure classification. By leveraging 12 objective MRI biomarkers—including global and regional metrics of mean signal intensity and standard deviation—the method enables robust quantitative assessment, significantly enhancing detection accuracy and inter-study reproducibility.	MRI
Nikpasand M.et al. 57	2024	CNN-based automated grading model	Automated grading of degeneration severity for lumbar discs and facet joints	This study introduces a dual-task unified framework based on a CNN that simultaneously performs Pfirrmann grading of intervertebral discs and Fujiwara grading of facet joints from lumbar MRI. The model incorporates optimized multi-scale input specifications (64×64 and 32×32 pixels) to accommodate anatomical size variations between structures, thereby enhancing recognition performance across different spinal components.	MRI
Niemeyer F.et al. 58	2024	DL model improved based on VGG-16 architecture	Automated grading of IDD severity in canine lumbar spine T2-weighted sagittal MRI images	This work presents the first DL-based system for automated grading of canine lumbar disc degeneration. The approach adapts human-designed architectures through structural modifications tailored to canine-specific anatomical and imaging characteristics, demonstrating successful cross-species model transfer and optimization.	MRI
Ke B.et al. 48	2024	DL model based on PP-YOLOv2	Automatically identify and classify three conditions from lumbar MRI images: normal disc, LDH, and spondylolisthesis.	This system adapts the PP-YOLOv2 object detection framework for lumbar MRI analysis, effectively addressing domain-specific challenges including small anatomical target localization and complex soft-tissue background interference. The optimized model demonstrates diagnostic accuracy significantly surpassing expert-level performance.	MRI
Liu Z.et al. 20	2024	ResNet-34	Automated identification of CLDH through lumbar sagittal MRI images	This approach enhances feature extraction in small-sample medical imaging by incorporating residual blocks and skip connections, effectively mitigating gradient vanishing issues and improving accuracy in the identification of calcified LDH.	MRI
Li Y.et al. 53	2025	YOLOv7-WRN-SVM	Achieving staging and classification of IDD based on T1ρ-MR from conventional T1-MR images.	This method employs a three-stage analytical pipeline: YOLOv7 for high-precision intervertebral disc localization, a Wide Residual Network for feature extraction and T1ρ value prediction, and a Support Vector Machine for degeneration staging. The complementary advantages of these architectures are systematically integrated. Additionally, Synthetic Minority Over-sampling Technique is applied to address class imbalance, substantially enhancing overall model robustness and performance.	MRI
Yilihamu E E.et al. 54	2025	Dual-branch multi-stage DL model based on YOLOv8 series	Achieves fast and accurate automated quantification and classification of LDH in lumbar sagittal T2-weighted MRI.	This framework introduces a dual-branch multi-stage workflow that integrates YOLOv8-based object detection, semantic segmentation, and key point detection models. It enables collaborative localization of intervertebral discs, pixel-level segmentation, and anatomical key point identification within the spinal canal region.	MRI
Nakamoto I.et al. 55	2025	WDRIV-Net	Automated hierarchical classification of single-type and combined-type lumbar disc degeneration in MRI.	This study introduces a weighted ensemble framework integrating four pre-trained models—DenseNet169, ResNet101, InceptionV3, and VGG19—to enhance classification performance in lumbar disc degeneration analysis. The approach outperforms both equal-weight ensemble strategies and individual models. It represents the first method specifically designed for automatic discrimination between single-type and combined-type lumbar disc degeneration, effectively addressing the limitations of conventional techniques in diagnosing complex degenerative conditions.	MRI
Guo Y.et al. 49	2025	GE-YOLOv8	Automatically detect and grade LDH based on lumbar axial T2-weighted MRI images.	This work introduces two novel components into the YOLOv8 architecture: a GS module that employs gradient splitting to reduce parameter volume by 21% and computational complexity while maintaining feature representation capacity—effectively addressing the high cost of the original C2f module—and an ECA mechanism that recalibrates feature channel weights to improve sensitivity in detecting small lesions.	MRI

In conclusion, AI technologies—through diverse and increasingly sophisticated model architectures—now support end-to-end assessment of disc degeneration, spanning basic detection fine-grained grading, cross-sequence and cross-species generalization and ensemble-driven performance gains. These advances significantly enhance the accuracy, efficiency, and standardization of MRI-based diagnostic workflows, laying a robust foundation for future clinical integration.

5.1 Limitations in sample size and data quality

Many studies rely on datasets that are limited in scale and derived from single institutions,^59,60 resulting in small or homogeneous samples that do not represent the general population. This increases the risk of overfitting and constraining model generalizability. A small-sample study demonstrated that while the model achieved an accuracy of 80% on the internal test set, its accuracy dropped to 63.23% on the external validation set. ²⁰ In addition, incomplete or biased patient data collection further restricts the spectrum of pathologies that models can accurately recognize. ⁶¹ In classification and grading tasks, Pfirrmann grade 1 constituted only 3.9% of the baseline sample size. The model’s accuracy in correctly classifying this grade was 68.3%, which is over 20 percentage points lower than the 92.6% accuracy achieved for grade 2, which comprised 64.9% of the initial sample. ⁶² Additionally, many studies fail to account for inter- and intra-observer variability during manual annotation. This compromises label consistency, embeds biases into the model, and ultimately undermines the reliability of both model training and evaluation.

5.2 Constraints in model performance

Several technical shortcomings remain evident in current AI applications. Certain segmentation methods exhibit tendencies toward oversegmentation—particularly in critical regions such as the thecal sac and intervertebral discs. ⁶³ In challenging scenarios like spinal stenosis, discontinuous or incomplete segmentation of the thecal sac is frequently observed, reflecting limited robustness in complex anatomical contexts. Some studies also employ outdated algorithm versions or omit state-of-the-art auxiliary techniques (e.g., iterative SAM architectures or advanced attention mechanisms), thereby limiting overall model capability. ⁶⁴ Performance inconsistency across studies ⁶⁵ raises concerns regarding diagnostic reliability, as erroneous predictions could lead to misdiagnosis or delayed treatment.

5.3 Insufficient validation and generalizability

The evaluation protocols adopted in many studies lack rigor and comprehensiveness. A limited number of studies have employed K-fold cross-validation. ⁴³ This method divides the dataset into K folds, iteratively using K-1 folds for training and the remaining fold for testing, repeating the process K times. The final performance metrics (e.g., accuracy, AUC) are averaged over the K test results. This approach yields more stable and reproducible outcomes than a single random train/test split, providing a better estimate of a model’s true generalization capability. Despite its proven efficacy, it is still frequently overlooked in many studies, ⁶⁶ leading to insufficient stability in performance evaluatio. More critically, many models are not validated on external datasets, ⁶⁷ many studies have shown that models demonstrate substantial performance gaps between internal and external datasets (e.g., GE-YOLOv8 achieved an mAP50 of 78% on the internal validation set versus 62.9% on the external test set). ⁴⁹ Furthermore, performance varies across different imaging protocols, ⁴ for instance, the model’s classification accuracy on spiral CT was 89.5%, over ten percentage points higher than on axial CT. This illustrates that the lack of external data validation can lead to overly optimistic performance metrics and severely compromise the model’s generalizability across different populations or imaging protocols. The absence of multi-center validation further exacerbates these issues. This is because the models fail to establish their applicability to real-world data derived from diverse institutions and various devices. Conversely, studies that have undergone rigorous multi-center validation ⁵⁴ can provide robust evidence of their model’s generalization capability, which constitutes an indispensable validation step when planning the clinical deployment of an AI model.

5.4 Barriers to clinical translation and implementation

Even with algorithmically superior models, the practical deployment of AI systems in clinical environments still faces multiple challenges.

Firstly, computational resources and deployment costs pose a substantial real-world barrier. Complex learning models often place high demands on computational resources. ⁴⁰ The significant computational requirements of advanced models such as Vision Transformers can lead to high hardware costs, creating a practical obstacle for resource-limited clinical institutions aiming to deploy AI systems. ⁴⁵ Secondly, integration with existing clinical workflows remains inadequate. Most research prototypes are not designed as software capable of seamless integration with hospital systems, and models often operate as isolated systems. ⁵¹ This inevitably increases the operational burden and time costs for clinical staff. Moreover, The “black-box” nature of many DL models ⁶⁸ remains a significant barrier to clinical integration. Without clear explanations for AI-driven decisions, clinicians may be hesitant to adopt such tools in high-stakes diagnostic settings. The lack of interpretability not affects trust but also complicates compliance with emerging regulatory standards for AI-based medical devices.

In conclusion, while AI shows transformative potential in disc disease imaging, overcoming these limitations—through improved data curation, model design, validation practices, and explainability—is essential to achieving clinically reliable and widely applicable diagnostic systems.