Back to all papers

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

November 20, 2026pubmed logopapers

Authors

Shimizu T,Kusunoki K,Nishida Y,Che X,Nakamura S,Arita K,Ohashi Y,Sugawara Y,Ishizu H,Yokota S,Yamada K,Iwasaki N

Affiliations (4)

  • Department of Orthopaedic Surgery, Faculty of Medicine and Graduate School of Medicine, Hokkaido University, Kita-15, Nishi-7, Kita-Ku, Sapporo, 060-8638, Japan. [email protected].
  • Department of Orthopaedic Surgery, Faculty of Medicine and Graduate School of Medicine, Hokkaido University, Kita-15, Nishi-7, Kita-Ku, Sapporo, 060-8638, Japan.
  • Department of Orthopaedic Surgery, Sapporo Higashi Tokushukai Hospital, Kita-33, Higashi-14, Higashi-Ku, Sapporo, 065-0033, Japan.
  • Future Corporation, Art Village Osaki Central Tower, 1-2-2 Osaki, Shinagawa-Ku, Tokyo, 141-0032, Japan.

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 from two vendors. A ResNet-50-based convolutional neural network was used to perform fracture region segmentation, fracture classification, pelvic fracture detection, and fracture line angle estimation. The diagnostic performance and fracture-line angle agreement were evaluated using internal and external validation datasets. An exploratory reader study involving six readers was conducted to assess the effects of CAD assistance. Overall classification accuracy was 95.7% (κ = 0.944) internally and 91.5% (κ = 0.887) externally; F1-score ranges were 0.927-0.982 and 0.667-0.952, respectively. The fracture-line angle mean absolute errors were 4.54° and 6.26°, approaching the interobserver errors between orthopedic specialists (3.64° and 5.18°), with corresponding ICCs of 0.817 and 0.736, respectively. In the exploratory reader study, residents showed higher sensitivity with CAD assistance for non-displaced femoral neck fractures (0.78 vs. 0.47), with high specificity. This multicenter CT-based CAD system demonstrated accurate and generalizable fracture region segmentation, proximal femoral fracture classification, and objective fracture line angle quantification across independent datasets. The proposed approach has the potential to support diagnosis and preoperative assessment of proximal femoral fractures.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.