fidoptBUU-LSPINE lumbar radiographs
Lateral radiographs of the lumbar spine of two patients of Burapha University Hospital, with the anterior and posterior corners of the endplates of L1 to L5 and of the top of S1 placed by physicians on each (BUU-LSPINE).
Read before comparing (7)
- Pool. 2,000 patients of Burapha University Hospital, Thailand, 1,318 women, aged 6 to 95, each with an AP and a lateral lumbar radiograph; only the lateral view is used. The release is the directly linked part of the 3,600-patient dataset; its 400-patient sample holds the same images and corners, with other diagnosis codes.
- Images. The provider's JPEGs, 1,096 to 3,040 by 2,128 to 3,037 pixels on the test pairs, 8-bit grey, stored as one component or as three equal ones, both within one pair, side markers and tags kept. The provider flipped every right-facing view to face left. No pixel size is given, so the unit is the pixel; vertebral bodies measure 118 to 250 pixels high, and the two images of a test pair differ in scale by a median factor of 1.10, at most 1.65. No mask.
- Truth. 22 points an image: the anterior and posterior corners of both endplates of L1 to L5 and of the upper endplate of S1, placed by the hospital's physicians and agreed at a review meeting of orthopaedists, radiologists and image analysts. No agreement between readers is published. Patients with a transitional lumbosacral vertebra keep the labels the physicians gave: their 'L5' or 'S1' is another bone than a typical patient's.
- Alignment. Each image keeps its own grid at the origin; the moving image's header carries the translation that puts the two images' centres together. elastix mutual information was tried first: periodic vertebrae drew it a level off on many pairs and four of 30 test pairs started over 900 pixels apart, against 623 at most from the centres.
- Baseline. With no registration the 30 test pairs start at a TRE of 285 pixels (78 to 623), rTRE 0.091, a 95th percentile of 367 pixels on average.
- Coverage. The points lie on the vertebral bodies of L1 to S1: nothing scores the posterior elements, the pelvis or the soft tissue.
- Overlap. No other dataset of the corpus holds Burapha images; patient numbers are unique in the release, and the paper says each record is a different patient.
No method has a result on every pair of this dataset yet.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 30 test pairs drawn with seed 2026 among the 1,000 pairs, and 30 development pairs, the last of the same seeded order. The release's 2,000 lateral images, one a patient, are put in the order of sha256("2026:BUULSpine:<number>") and paired two by two, the first fixed. One hospital: the draw is not stratified.
- Truth
- landmarks by experts the dataset's 22 corners on each lateral image: the anterior and posterior corners of the upper and lower endplates of L1 to L5 and of the upper endplate of S1, annotated by the hospital's physicians (radiologists and orthopaedic doctors) and verified at a review meeting (Klinwichit et al. 2023)
- Source
- services.informatics.buu.ac.th/spine
doi.org/10.3390/app13158646
drive.google.com/file/d/1qFdU420eLo6rSR0lyuWB9M3rfknXFhvm/view - Licence
- Burapha University Faculty of Informatics end-user licence agreement: no re-identification, acknowledge and cite, no endorsement; redistribution not addressed
30 pairs for tuning, never tested
1576_0766, 1469_1419, 1252_2255, 0722_1354, 0170_0710, 2274_1143, 0777_1299, 2252_2238, 1335_0984, 1901_1486, 0161_2134, 2468_2472, 1561_2381, 1227_0035, 1623_1793, 0756_1182, 0851_0264, 0972_2041, 2571_1665, 1703_1007, 0467_2592, 0244_0892, 0749_0776, 1537_2408, 0857_0377, 2240_2612, 0459_2462, 0458_0246, 0820_0940, 0828_2226