68/100

Moderate evidence

Cite
More
EmailLinkedInX

AI reading hip CT scans got 91.5% of fracture types right in outside hospitals, study finds

Study title: Multicenter external validation of CT-based deep learning for proximal femoral fracture classification and quantitative fracture-line assessment

Evidence type
Controlled or validated comparison
Sample
n = 550
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Multicenter external validation of CT-based deep learning for proximal femoral fracture classification and quantitative fracture-line assessment

Researchers built a deep learning system that reads coronal CT images to classify proximal femoral (upper thigh bone) fractures. It reached 95.7% accuracy on internal validation and 91.5% on 200 patients from independent institutions.

Read the study abstract

To develop and externally validate a CT-based deep learning CAD system for proximal femoral fracture classification and fracture-line angle quantification and to preliminarily evaluate its clinical utility in a reader study. In this multicenter retrospective study, coronal CT images from four institutions were used for training (259 patients, 19,844 annotated images) and internal validation (91 patients, 92 hips). External validation included 200 patients from independent institutions scanned with CT systems fro…

Multicenter external validation of CT-based deep learning for proximal femoral fracture classification and quantitative fracture-line assessment
"Medical Records" by Georgia National Guard, CC BY 2.0.

The gain

In a multicenter study with 200 externally validated patients, a CT deep learning system classified proximal femoral fractures with 91.5% accuracy, and residents' sensitivity for non-displaced femoral neck fractures rose from 0.47 to 0.78 with its help.

In brief

The system also measured the fracture-line angle with a mean absolute error of 6.26° externally, close to the 5.18° difference seen between orthopedic specialists. In an exploratory six-reader study, residents detected more non-displaced femoral neck fractures with the tool's help, with sensitivity rising from 0.47 to 0.78.

Performance dropped outside the development hospitals, and F1-scores for some fracture categories fell as low as 0.667 externally. The reader study was exploratory, so the clinical benefit is preliminary.

Main points

  1. Training used 259 patients and 19,844 annotated coronal CT images from four institutions; internal validation used 91 patients (92 hips).
  2. External validation covered 200 patients from independent institutions scanned on CT systems from two vendors.
  3. Classification accuracy was 95.7% (κ = 0.944) internally and 91.5% (κ = 0.887) externally.
  4. F1-score ranges were 0.927-0.982 internally and 0.667-0.952 externally.
  5. Fracture-line angle errors were 4.54° and 6.26°, versus 3.64° and 5.18° between orthopedic specialists.
  6. Six readers took part in an exploratory study; residents' sensitivity for non-displaced femoral neck fractures rose from 0.47 to 0.78 with high specificity.

The full story

The study was retrospective and multicenter. It used a ResNet-50-based convolutional neural network to segment fracture regions, classify fractures, detect pelvic fractures and estimate the fracture-line angle on coronal CT images.

Agreement on fracture-line angles was moderate to good, with intraclass correlation coefficients of 0.817 internally and 0.736 externally. The angle errors approached the variation between orthopedic specialists.

The authors conclude the approach could support diagnosis and preoperative assessment of proximal femoral fractures. The reader study involved six readers and was described as exploratory.

Sources

1 on file
  1. Skeletal Radiologydoi:10.1007/s00256-026-05447-1
    Peer-reviewed2026-11-20

Talk it over

Three questions to start on, two more if you have time
Accuracy was 95.7% internally and 91.5% on 200 patients from independent institutions, with some F1-scores as low as 0.667. How would you weigh that drop against the gain for residents?

Start with a question

AI reading hip CT scans got 91.5% of fracture types right in outside hospitals, study findshttps://truvace.com/item/multicenter-external-validation-of-ct-based-deep-learning-for-proximal-femoral-f

  1. Accuracy was 95.7% internally and 91.5% on 200 patients from independent institutions, with some F1-scores as low as 0.667. How would you weigh that drop against the gain for residents?

    Answer in the thread
  2. Residents' sensitivity for non-displaced femoral neck fractures rose from 0.47 to 0.78 in an exploratory study with six readers. What can so few readers tell us, and what result would settle whether it helps patients?

    Answer in the thread
  3. Which patients or hospitals might be missing from a study scanned on CT systems from two vendors, and how could that change the 91.5% result?

    Answer in the thread
If you have more time
  • Look up how fracture-line angles are used in preoperative planning. How much does an error of 6.26° matter compared with the 5.18° between orthopedic specialists?

    Answer in the thread
  • If a hospital you know of adopted this system, who should check its output on non-displaced fractures, and how would they do it?

    Answer in the thread

What readers say