Spondylolisthesis Vertebral Landmark

The dataset used in this study consists of 716 sagittal lumbar spine X-ray images:

208 images from a proprietary dataset of Honduran patients diagnosed with spondylolisthesis.
508 images from the publicly available BUU-LSPINE dataset, filtered for sagittal views only.

Each image was manually annotated to identify vertebral landmarks (corners of each vertebral body from L3 to S1). These annotations include:
Bounding boxes for vertebrae.
Four anatomical corner keypoints per vertebra.
Annotations were formatted in JSON formats compatible with PyTorch's Keypoint R-CNN.

The dataset was split as follows:
69% (494 images) for training.
29% (206 images) for validation.
16 images reserved for clinical evaluation by experts.

The curated dataset of 698 annotated images can serve as a public benchmark for future research in vertebra detection, spinal alignment analysis, or degenerative disorder classification.

The consistent performance across variable image qualities and patient anatomies enhances its potential for deployment in real-world radiology workflows.