Table 1 Characteristics of the development and test datasets.

From: Assessment of a deep-learning system for fracture detection in musculoskeletal radiographs

 

Development dataset

Test dataset

Radiographs

 No. of hospitalsa

15

15

 No. of radiographs

715,343

16,019

 No. of radiographic views

16b

9c

 No. of anatomical regions

16

16

 Median (range) radiographs per anatomical region

40,658 (6249–106,705)

1000 (774–1079)

 No. of radiographs with fracture(s) (%)

82,830 (12%)

2415 (15%)

 No. of fracture bounding-box annotations

97,559

2718d

 No. of bounding-box annotations per fractured radiograph, mean (range)

1.2 (1–13)

1.1 (1–6)

Patients

 No. of patients

314,866

12,746

 Median (range) patients per anatomical region

18,952 (3022–71,484)

909 (326–1042)

 No. of male (%)

137,929e (44%)

5520 (43%)

 Median (range) patient age in years

54 (0–90)f

55 (22–90)

Annotators

 No. of orthopedic surgeons (median years experience post-residency)

18 (16)

11 (13)

 No. of radiologists (median years experience post-residency)

11 (13)

7 (13)

  1. No radiographs used for testing were in the development dataset.
  2. aDatasets sampled from the MedStar Health System located in Baltimore, MD, Washington, D.C., Olney, MD, Leonardtown, MD, and Clinton, MD, the CarePoint Health System in Bayonne, NJ, Jersey City, NJ, and Hoboken, NJ as well as the Hospital for Special Surgery (HSS) in New York, NY and Orthopedic Institute for Children in CA.
  3. bNumber of unique radiographic views estimated through a manual review of 20,000 randomly sampled radiographs across anatomical regions.
  4. cViews were collapsed for statistical analyses into frontal view (frontal; frontal dorso-plantar; frontal inlet-outlet), lateral view (axillary; frog-leg lateral; lateral; y), and oblique view (oblique; oblique-mortise).
  5. d2718 reflects unique fracture sites after fusing the 3 reference standard annotations per image through a pixel-wise majority vote.
  6. e602 patients were missing biological sex information.
  7. fPatient age missing for 43% of the development dataset because patient age was removed from radiographs collected at HSS. De-identification procedures capped patient age at 90 years. In the development dataset, 0.1% of radiographs were from patients 0 to 10 years of age, and 2.95% were from patients 10 to 20 years of age. By design, no radiographs in the test dataset were from patients <22 years of age.