Deep Learning to Differentiate Benign and Malignant Vertebral Fractures at Multidetector CT
Sarah C Foreman1, David Schinz1, Malek El Husseini1
1From the Departments of Radiology (S.C.F., A.S.D., G.C.F., M.R.M.) and Neuroradiology (D.S., M.E.H., M.R., M.C.M., B.W., B.J.S., J.S.K.), Klinikum Rechts der Isar, Technische Universität München, Ismaninger Strasse 22, 81675 Munich, Germany; Departments of Radiology (S.S.G., J.W.) and Neuroradiology (R.S., A.S.G.), University Hospital Munich (LMU), Munich, Germany; and German Cancer Consortium (DKTK), Partner Site Munich, and German Cancer Research Center (DKFZ), Heidelberg, Germany (B.W.).
Deep learning models accurately differentiate benign from malignant vertebral fractures, matching expert radiologist performance. These AI tools show high potential for improving diagnostic accuracy in challenging cases.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Distinguishing benign from malignant vertebral fractures is diagnostically challenging.
- Computed tomography (CT) is a key imaging modality for evaluating vertebral fractures.
- Deep learning (DL) offers potential for automated analysis of medical images.
Purpose of the Study:
- To assess the reliability of CT-based deep learning models in differentiating benign from malignant vertebral fractures.
- To compare the diagnostic performance of DL models against human expert readers.
Main Methods:
- Retrospective analysis of CT scans from patients with confirmed benign or malignant vertebral fractures.
- Development of a 3D U-Net encoder-classifier deep learning architecture with data augmentation.
- Evaluation using area under the receiver operating characteristic curve (AUC) on internal and external test sets.
- Comparison of model performance with radiology residents and a fellowship-trained radiologist.
Main Results:
- The best-performing DL model achieved AUCs of 0.85 (internal) and 0.75 (external).
- Incorporating an uncertainty category improved performance to AUCs of 0.91 (internal) and 0.76 (external).
- DL models outperformed radiology residents (AUCs 0.69-0.71) and matched fellowship-trained radiologist performance (AUCs 0.71-0.86).
Conclusions:
- Developed CT-based deep learning models demonstrate high discriminatory power for vertebral fracture characterization.
- These AI models show potential to assist radiologists, matching or exceeding the performance of less experienced readers.
- The findings support the clinical utility of AI in improving the diagnosis of malignant vertebral lesions.
