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 for automatically detecting and classifying degenerative conditions in lumbar spine discs and vertebrae from 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 with significantly reduced annotation effort compared to existing methods.

What Problem Does This Paper Address?

Spinal degeneration is a widespread issue affecting many individuals, leading to pain and mobility problems. Accurate diagnosis is crucial for effective treatment, but current methods rely on laborious manual analysis of medical images like MRIs by clinical experts. While automated systems exist, they often follow a two-stage process: first localizing anatomical structures (discs and vertebrae) and then classifying their condition. This can be inefficient and requires extensive expert annotation. This paper aims to streamline this process by proposing a one-stage framework that simultaneously localizes and classifies these structures, reducing the burden on experts and improving diagnostic efficiency.

Methods

SpineOne is built upon three core novel techniques integrated into a one-stage detection framework:

  • One-Channel-Per-Class (OCPC) Keypoint Heatmap: Instead of traditional heatmaps that might use multiple channels per keypoint or class, OCPC uses a single channel for each class (normal vs. degenerative) to represent all keypoints of that class. This design is compact and leverages the spatial separation of discs and vertebrae to capture correlations between keypoints, facilitating simultaneous localization and classification.
  • Dual Self-Attention Modules (PAM & CAM): To better distinguish between the similar visual features of discs and vertebrae, SpineOne employs two attention modules:
  • Position Attention Module (PAM): Aggregates features across all positions in an image, capturing long-range spatial dependencies.
  • Channel Attention Module (CAM): Integrates features across different channels, allowing the network to focus on more informative feature maps.
  • Gradient-Guided Objective Association (OA): This mechanism addresses the fact that heatmap generation (for localization and classification) is more critical than offset regression (for fine-tuning localization). OA uses the gradient of the heatmap loss with respect to the heatmap itself to guide the learning of the offset, making the heatmap learning more effective and improving overall performance.

Datasets

The primary dataset used in this paper is the Spinal Disease Intelligent Diagnosis Tianchi Competition (SDID-TC) dataset. This dataset consists of 550 lumbar spine MRI exams. The authors randomly split this dataset into 400 exams for training and 150 for testing. They specifically utilized the sagittal T2-weighted MRI sequences, as these sequences highlight cerebrospinal fluid, aiding in the visualization of discs and vertebrae and their surrounding structures.

Key Insights and Ideas

  • One-Stage Efficiency: The paper demonstrates that a one-stage approach can be highly effective for this complex task, outperforming traditional two-stage methods by leveraging the intrinsic correlations between localization and classification.
  • Annotation Reduction: By focusing on keypoint centroids and classification rather than full segmentation or bounding boxes, the annotation effort required from clinical experts is significantly reduced, making the approach more practical.
  • Attention for Discrimination: The dual attention modules are crucial for differentiating between discs and vertebrae, which share high inter-structural similarity. The paper highlights that position attention is particularly impactful.
  • Gradient Guidance for Learning: The OA mechanism is a clever way to ensure that the more critical task of heatmap generation (localization and classification) effectively guides the learning of the less critical offset regression, leading to improved overall performance.

Key Results

  • SpineOne achieved state-of-the-art performance on the SDID-TC dataset, outperforming existing one-stage and two-stage methods by a significant margin. For instance, it achieved a macro mean F1 score of 0.859 for discs and 0.849 for vertebrae when using the DeepLabv3+(Res101) backbone, surpassing other methods by 3%-6.4% in F1 score.
  • The framework is computationally efficient, with an inference time of 0.029 seconds, significantly faster than two-stage methods which can take over 0.3 seconds.
  • Ablation studies showed that each of the proposed components (OCPC heatmap, attention modules, OA mechanism) contributed to performance improvements, with the OCPC heatmap being particularly essential.
  • The framework also demonstrated competitive results on a public dataset for vertebrae localization and identification (MICCAI 2014 challenge), even when adapted for 3D CT volumes.

Limitations and Future Work

  • Dataset Volume: The authors acknowledge that augmenting the volume of the SDID-TC dataset would be beneficial for further improving performance.
  • Degenerative Grading: The current framework classifies degeneration into binary categories (normal/degenerative). Future work could extend this to a more detailed grading system for discs and vertebrae.
  • Generalizability: While the framework shows promise for other medical diagnosis applications, further verification of the proposed techniques in different medical scenarios is suggested.
  • Slices Selection: While the middle sagittal T2 slices are optimal, the paper notes that performance can be affected by the spacing and resolution of input slices, particularly in low-resolution cases.
  • Offset Importance: The paper notes that the offset learning is less critical than heatmap learning, and models without offset outputs still perform competitively, albeit slightly lower.

Summary updated

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},
}

Loading related content...

Figure