Schaapman, Okke (2025) Performing Independent External Validation of a Convolutional Neural Network (CNN) for the Detection of Suspected Scaphoid Fractures: Trying to Tackle an Important Clinical Diagnostic Challenge. thesis, Medicine.
Full text available on request.Abstract
One in every five scaphoid fractures is missed on initial radiographical examination, prompting the need for advanced imaging techniques and increasing the risk of overdiagnosis. To tackle this diagnostic challenge, artificial intelligence (AI) algorithms, particularly convolutional neural networks (CNNs), are being explored for their potential in fracture detection. While CNNs have demonstrated promising results in identifying both apparent and occult fractures, their generalizability remains under-researched. Materials and Methods This retrospective cohort study included 415 patients who presented with a suspected scaphoid fracture. All patients underwent advanced imaging to establish an accurate reference standard. Four scaphoid projections (posteroanterior, uptilt, lateral and oblique) were presented per patient to one CNN for analysis. Sensitivity, specificity, accuracy and area under the receiver operating characteristic curve (AUC) were calculated to quantify performance. Results 176 patients had a fracture, and 239 patients did not have a fracture. 34 fractures were only visible on advanced imaging (occult). The CNN demonstrated a sensitivity of 76.70% (95% confidence interval (CI) [69.93–82.34%]), specificity of 88.51% (95% CI [83.80–91.98%]), accuracy of 83.45% (95% CI [79.56–86.73%]) and an AUC of 0.890 (95% CI [0.856–0.924]). The algorithm identified 8 out of 34 occult fractures (sensitivity of 23.53% (95% CI [12.44– 40.00%])). Conclusions This study was the first performing independent external validation of a scaphoid fracture detection CNN. While the standalone application of scaphoid fracture detection CNNs show considerable potential, CNNs remain subject to important limitations. These findings underscore the necessity of independent external validation prior to clinical implementation. Keywords Scaphoid fracture, Medical imaging, Artificial intelligence, Machine learning, Convolutional neural network, External validation
| Item Type: | Thesis (UNSPECIFIED) |
|---|---|
| Supervisor name: | Jaarsma, Professor Dr. Ruurd L. and Doornberg, Professor Dr. Job N. |
| Faculty: | Medical Sciences |
| Date Deposited: | 01 May 2026 13:04 |
| Last Modified: | 01 May 2026 13:04 |
| URI: | https://umcg.studenttheses.ub.rug.nl/id/eprint/3939 |
Actions (login required)
 |
View Item |