(P 176) DEEP LEARNING MODEL FOR DIFFERENTIATING BETWEEN NEOPLASTIC PATHOLOGIC FRACTURE AND NONPATHOLOGIC FRACTURE FROM PLAIN HIP RADIOGRAPHS

Objective: Although radiographs are the first-line imaging modality, differentiating between neoplastic pathologic fractures and nonpathologic fractures on radiographs can sometimes be challenging. Misdiagnosing a neoplastic fracture as nonpathologic and providing standard fracture treatment may result in devastating outcomes, including limb amputation or death. This study aimed to develop and evaluate a deep learning model capable of distinguishing neoplastic pathologic fractures from nonpathologic fractures on hip radiographs to enhance diagnostic accuracy.

Methods: This retrospective, multicenter study analyzed anteroposterior hip radiographs from patients who visited the emergency department at four distinct institutions. The deep learning model was trained and tested on data from 338 patients at a single institution and externally validated on data from 488 patients across three additional institutions.

Results: The model achieved an overall accuracy of 88.0%, with a sensitivity of 88.2% and a specificity of 87.9% on the internal test set. It was then externally tested on 488 patients from institutions separate from where it was developed, including 67 cases of neoplastic pathologic fractures and 421 nonpathologic fractures. The model achieved an overall accuracy of 84.8%, sensitivity of 91.0%, and specificity of 78.6%. Its performance was comparable to that of general orthopedic surgeons.

Conclusion: Our deep learning model demonstrates robust performance in distinguishing neoplastic pathologic fractures from nonpathologic fractures on hip radiographs, offering a clinically meaningful diagnostic aid. By enabling early and accurate identification of neoplastic lesions, especially in resource-constrained settings, the tool has the potential to significantly reduce misdiagnosis-related morbidity. To enhance clinical accessibility and impact, the model is freely available at https://pathfxdx.org.

Performance of the model compared to experts