Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spinal Osteophyte Detection via Robust Patch Extraction on minimally annotated X-rays

Soumya Snigdha Kundu, Yuanhan Mo, Nicharee Srikijkasemwat, Bartłomiej W. Papiez

TL;DR

This paper presents one of the first efforts towards automated spinal osteophyte detection in spinal X-rays based on deep learning-driven vertebrae segmentation and the enlargement of mask contours, which demonstrates that even with limited annotations, Seg-Patch can deliver superior performance for detection of tiny structures such as osteophytes.

Abstract

The development and progression of arthritis is strongly associated with osteophytes, which are small and elusive bone growths. This paper presents one of the first efforts towards automated spinal osteophyte detection in spinal X-rays. A novel automated patch extraction process, called SegPatch, has been proposed based on deep learning-driven vertebrae segmentation and the enlargement of mask contours. A final patch classification accuracy of 84.5\% is secured, surpassing a baseline tiling-based patch generation technique by 9.5%. This demonstrates that even with limited annotations, SegPatch can deliver superior performance for detection of tiny structures such as osteophytes. The proposed approach has potential to assist clinicians in expediting the process of manually identifying osteophytes in spinal X-ray.

Spinal Osteophyte Detection via Robust Patch Extraction on minimally annotated X-rays

TL;DR

This paper presents one of the first efforts towards automated spinal osteophyte detection in spinal X-rays based on deep learning-driven vertebrae segmentation and the enlargement of mask contours, which demonstrates that even with limited annotations, Seg-Patch can deliver superior performance for detection of tiny structures such as osteophytes.

Abstract

The development and progression of arthritis is strongly associated with osteophytes, which are small and elusive bone growths. This paper presents one of the first efforts towards automated spinal osteophyte detection in spinal X-rays. A novel automated patch extraction process, called SegPatch, has been proposed based on deep learning-driven vertebrae segmentation and the enlargement of mask contours. A final patch classification accuracy of 84.5\% is secured, surpassing a baseline tiling-based patch generation technique by 9.5%. This demonstrates that even with limited annotations, SegPatch can deliver superior performance for detection of tiny structures such as osteophytes. The proposed approach has potential to assist clinicians in expediting the process of manually identifying osteophytes in spinal X-ray.
Paper Structure (14 sections, 5 figures, 1 table)

This paper contains 14 sections, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Data
  3. Methodologies
  4. Robust Patch Creation
  5. Tiling
  6. SegPatch: Segmentation-driven Patch Extraction
  7. Classification of Osteophytes
  8. Results
  9. Discussion
  10. Solving the class imbalance in Tiling
  11. Problems with generating more precise patches
  12. Patch size variability in SegPatch
  13. Poor Performance from off-the-shelf detectors
  14. Conclusion

Figures (5)

  • Figure 1: Example of annotations for cervical (left) and lumbar (right) X-ray scan available from NHANESII$^2$ data set.
  • Figure 2: A representation of the SegPatch pipeline with a resulting patch. The green box highlights the necessity to expand contours to fully encompass the osteophytes within the boundary. Red dots denote the osteophytes.
  • Figure 3: Saliency results for a positive and a negative patch. Red dots: Osteophytes. Purple boxes: Contours.
  • Figure 4: Visualisation of poor performance from off-the-shelf detector - FasterRCNN. Purple circles denote annotation location and squares denote predicted bounding box.
  • Figure 5: Example of difficult (but successful) to label scans. (Left): High curvature of spine. (Right): Presence of artifact (text) denoted by red arrow with with no osteophyte labels.