SpineOne: A One-Stage Detection Framework for Degenerative Discs and Vertebrae

Jiabo He, Wei Liu, Yu Wang, Xingjun Ma, Xian-Sheng Hua

TL;DR

This paper introduces SpineOne, a novel one-stage deep learning framework that simultaneously localizes and classifies degenerative discs and vertebrae in spinal MRI scans. By integrating a new keypoint heatmap design, attention modules, and a gradient-guided objective association mechanism, SpineOne achieves state-of-the-art performance, significantly improving diagnostic efficiency for spinal degeneration.

What Problem Does This Paper Address?

Spinal degeneration is a widespread condition affecting many individuals, leading to pain and mobility issues. Accurate diagnosis is crucial for effective treatment, but the traditional process of identifying degenerative discs and vertebrae from medical images like MRIs is time-consuming and labor-intensive for clinical experts. Existing automated methods often employ a two-stage approach: first localizing anatomical structures and then classifying their condition. This paper aims to streamline this process by proposing a single, efficient framework that performs both localization and classification simultaneously, reducing the burden on experts and potentially leading to faster, more accurate diagnoses.

Methods

SpineOne is built upon three core innovations:

  1. One-Channel-Per-Class (OCPC) Keypoint Heatmap: Instead of traditional heatmaps that might use multiple channels for each keypoint and class, SpineOne uses a single channel per class. This design is advantageous because it compacts information, allowing the network to learn correlations between keypoints of the same class and facilitating simultaneous localization and classification. This approach is particularly useful for anatomical structures like discs and vertebrae that have a degree of spatial separation.
  2. Dual Self-Attention Modules (PAM & CAM): To better distinguish between the similar visual features of discs and vertebrae, SpineOne incorporates two attention modules:
  3. Position Attention Module (PAM): This module helps the network understand spatial relationships within the image by aggregating features based on their positions.
  4. Channel Attention Module (CAM): This module focuses on inter-channel correlations, allowing the network to weigh the importance of different feature maps. Together, these modules help the network learn more discriminative representations for discs and vertebrae.
  5. Gradient-Guided Objective Association (OA): This novel mechanism aims to improve the learning process by explicitly linking the heatmap (localization and classification) and offset (fine-tuning localization) learning objectives. It uses the gradient of the heatmap loss to guide the offset learning, making the heatmap learning more effective. This is particularly important as accurate classification is often more critical than precise localization in spinal diagnosis.

Datasets

The primary dataset used in this research is the Spinal Disease Intelligent Diagnosis Tianchi Competition (SDID-TC) dataset. This dataset comprises 550 spinal MRI exams. The authors randomly split this dataset into 400 exams for training and 150 for testing. They specifically utilize the sagittal T2-weighted sequences, as these are commonly used by clinicians for diagnosing degenerative spine conditions.

Key Insights and Ideas

  • One-Stage is Sufficient (and Better): The paper challenges the necessity of two-stage approaches for this task, demonstrating that a well-designed one-stage framework can outperform them. This is attributed to the ability of one-stage methods to leverage intrinsic correlations between different anatomical structures more effectively.
  • OCPC Heatmap's Simplicity and Power: The OCPC heatmap design is a key enabler, offering a compact and generic solution that simplifies the network architecture while improving performance. Its ability to capture geometrical and classification correlations is crucial for this specific task.
  • Attention for Discrimination: The dual attention modules are vital for differentiating between discs and vertebrae, which share visual similarities. The PAM's emphasis on spatial interaction appears to be particularly impactful.
  • Gradient Guidance for Objective Alignment: The OA mechanism is a clever way to ensure that the more critical heatmap learning task effectively guides the offset learning, leading to better overall performance without introducing additional hyperparameters.

Key Results

  • State-of-the-Art Performance: SpineOne significantly outperforms existing one-stage and two-stage methods on the SDID-TC dataset, achieving the highest F1 scores for both disc and vertebrae detection and classification. For example, with the DeepLabv3+(Res101) backbone, SpineOne achieved an F1 score of 0.859 for discs and 0.849 for vertebrae, outperforming other methods.
  • Component Contributions: Ablation studies show that each of SpineOne's proposed components (OCPC heatmap, PAM, CAM, OA) contributes to performance improvement. The integrated framework yields the best results.
  • Efficiency: SpineOne is significantly faster during inference (0.029 seconds) compared to two-stage methods (over 0.3 seconds), making it more practical for clinical applications.
  • Competitive on Public Benchmark: SpineOne also demonstrated strong performance on a public 3D CT dataset for vertebrae localization and identification, showcasing its generalizability.

Limitations and Future Work

  • Dataset Volume: The authors acknowledge that augmenting the volume of the SDID-TC dataset would be beneficial.
  • Degenerative Grading: The current work focuses on binary classification (normal/degenerative). Future work could explore more granular degenerative grading.
  • Generalizability: While promising, the authors suggest verifying the proposed techniques in other medical scenarios.
  • Offset Importance: The paper notes that offset learning is less critical than heatmap learning, and while the OA mechanism improves performance, the improvement from the offset itself is marginal.
  • Slices Selection: While 7 slices in the middle of the sagittal T2 sequence were found to be optimal, further exploration into optimal slice selection strategies could be beneficial.

Summary updated

Abstract

Spinal degeneration plagues many elders, office workers, and even the younger generations. Effective pharmic or surgical interventions can help relieve degenerative spine conditions. However, the traditional diagnosis procedure is often too laborious. Clinical experts need to localize discs and vertebrae as a preliminary step of pathological diagnosis. Machine learning systems have been developed to aid this procedure generally following a two-stage methodology: first perform anatomical localization, then pathological classification. Towards more efficient and accurate diagnosis, we propose a one-stage detection framework termed SpineOne to simultaneously localize and classify degenerative discs and vertebrae from magnetic resonance imaging (MRI) slices. SpineOne is built upon the following three key techniques: 1) a new design of the keypoint heatmap to facilitate simultaneous keypoint localization and classification; 2) the use of attention modules to better differentiate the representations between discs and vertebrae; and 3) a novel gradient-guided objective association mechanism to associate multiple learning objectives at the later training stage. Empirical results on the Spinal Disease Intelligent Diagnosis Tianchi Competition (SDID-TC) dataset of 550 exams demonstrate that our approach surpasses existing methods by a large margin.
BibTeX
@inproceedings{he2021spineone,
  author = {Jiabo He and Wei Liu and Yu Wang and Xingjun Ma and Xian-Sheng Hua},
  title = {{SpineOne: A One-Stage Detection Framework for Degenerative Discs and Vertebrae}},
  booktitle = {BIBM},
  year = {2021},
  doi = {10.1109/bibm52615.2021.9669541},
}
Figure