Critical Validation of Deep Learning-Based Cervical Vertebral Fracture Detection on Computed Tomography Localizer Radiographs.
Authors
Affiliations (11)
Affiliations (11)
- Institute of Databases and Information Systems, Ulm University, Ulm, Germany. [email protected].
- DASU - Transferzentrum für Digitalisierung, Analytics & Data Science Ulm, Ulm, Germany. [email protected].
- Institute for Artificial Intelligence, Ravensburg-Weingarten University of Applied Sciences, Weingarten, Germany.
- DASU - Transferzentrum für Digitalisierung, Analytics & Data Science Ulm, Ulm, Germany.
- Institute of Databases and Information Systems, Ulm University, Ulm, Germany.
- Department of Computer Science, Ulm University of Applied Sciences, Ulm, Germany.
- Department of Diagnostic and Interventional Radiology, University Hospital Ulm, Ulm, Germany.
- Department of Nuclear Medicine, University Hospital Ulm, Ulm, Germany.
- Department of Diagnostic and Interventional Neuroradiology, School of Medicine and Health, TUM Klinikum Rechts der Isar, Technical University of Munich, Munich, Germany.
- TUM-Neuroimaging Center, TUM Klinikum Rechts der Isar, Technical University of Munich, Munich, Germany.
- Department of Trauma-, Hand-, Plastic- and Reconstructive Surgery, University Hospital Ulm, Ulm, Germany.
Abstract
Localizer radiographs are low-resolution two-dimensional overview images acquired prior to computed tomography (CT) scans. Their automated analysis could support the early detection of cervical vertebral fractures in trauma patients and thus enable immediate adjustment of imaging protocols. This study investigated to what extent deep learning-based methods can reliably detect such fractures on localizer radiographs, with a focus on critically validating model behavior beyond standard performance metrics. On a custom-annotated dataset of 346 trauma patients, two approaches were compared: a full-image binary classification and a two-step pipeline that first localizes the cervical spine and then classifies the cropped region. Model validity was assessed through a permutation test, occlusion testing, gradient-weighted class activation mapping, an analysis of potential confounders, and a subgroup analysis by anatomical fracture location. The full-image classifier formally achieved the highest test-set performance with a precision-recall area under the curve (PR-AUC) of 0.477, compared with 0.349 for the cervical-spine-level classifier. Our validity analysis indicates that the full-image classifier does not primarily rely on the fracture region itself, rendering it unsuitable for clinical use. This work therefore presents a case study of shortcut learning and illustrates validation methods that may be of value for comparable studies on small datasets with visually subtle targets.