Page 101 - GTM-4-3

Global Translational Medicine CNNs for overfitting and generalizability in fracture detection

introduced potential selection biases and constrained the Ethics approval and consent to participate
assessment of model performance in prospective, real-
world applications, suggesting that further research is Not applicable.
necessary to refine the system for use as a screening tool Consent for publication
rather than a definitive diagnostic instrument.
Not applicable.
Collaborative learning across trauma networks
could enhance generalizability while adhering to data Availability of data
governance constraints, though this requires standardized
annotation frameworks to ensure cross-site label All data analyzed have been presented in the paper.
consistency. Ultimately, progression to clinical utility References
demands co-design with radiologists to align AI outputs
with interpretable diagnostic criteria. 1. Kutbi M. Artificial intelligence-based applications for bone
fracture detection using medical images: A systematic
5. Conclusion review. Diagnostics (Basel). 2024;14(17):1879.
While the model demonstrated high overall accuracy, doi: 10.3390/diagnostics14171879
the clinical implications of FPs and FNs warrant careful 2. Dankelman LHM, Schilstra S, IJpma FFA, et al. Artificial
consideration. FPs, where the model incorrectly identifies intelligence fracture recognition on computed tomography:
a fracture, can lead to unnecessary further investigations, Review of literature and recommendations. Eur J Trauma
increased patient anxiety, and potential delays in Emerg Surg. 2022;49:681-691.
appropriate treatment for the actual underlying condition. doi: 10.1007/s00068-022-02128-1
For example, an FP might trigger additional imaging
studies, such as computed tomography scans or magnetic 3. Sharma S. Artificial intelligence for fracture diagnosis
resonance imaging, which expose patients to radiation or in orthopedic X-rays: Current developments and future
contrast agents and add to healthcare costs. Conversely, potential. SICOT J. 2023;9:21.
FNs, where the model fails to identify a true fracture, pose doi: 10.1051/sicotj/2023018
a more serious risk. Missed fractures can result in delayed 4. Thomas D. How AI and convolutional neural networks can
or inadequate treatment, potentially leading to long-term revolutionize orthopaedic surgery. J Clin Orthop Trauma.
complications such as malunion, nonunion, or chronic 2023;40:102165.
pain. Moreover, in weight-bearing bones, a missed fracture
could lead to further injury and disability. Therefore, doi: 10.1016/j.jcot.2023.102165
while AI can be a valuable tool, clinicians should always 5. Lopez Pinaya WH, Vieira S, Garcia-Dias R, Mechelli A.
interpret the model’s output in conjunction with their own Convolutional neural networks. In: Machine Learning.
clinical judgment, patient history, and other diagnostic Netherlands: Elsevier; 2020. p. 173-191.
information to minimize the impact of both FPs and FNs. doi: 10.1016/b978-0-12-815739-8.00010-9
Acknowledgments 6. Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional
neural networks: An overview and application in radiology.
None. Insights Imaging. 2018;9:611-629.
Funding doi: 10.1007/s13244-018-0639-9
7. Ketkar N, Moolayil J. Convolutional neural networks.
None.
In: Deep Learning with Python. New York: Apress; 2021.
Conflict of interest p. 197-242.
doi: 10.1007/978-1-4842-5364-9_6
The authors declare they have no competing interests.
8. Kuo RYL, Harrison C, Curran TA, et al. Artificial intelligence
Author contributions in fracture detection: A systematic review and meta-analysis.
Radiology. 2022;304:50-62.
Conceptualization: All authors
Formal analysis: All authors doi: 10.1148/radiol.211785
Investigation: All authors 9. Jung J, Dai J, Liu B, Wu Q. Artificial intelligence in fracture
Methodology: All authors detection with different image modalities and data types:
Writing–original draft: All authors A systematic review and meta-analysis. PLOS Digit Health.
Writing–review & editing: All authors 2024;3:e0000438.

Volume 4 Issue 3 (2025) 93 doi: 10.36922/gtm.8526