Enhancing Fracture Detection in Remote Settings: Evaluating the Efficacy of FIXUS AI Deep Learning Algorithms in Identifying Fifth Metatarsal Fractures Using Mixed-Quality X-rays
Session Number
CMPS(ai) 10
Advisor(s)
Atta Taseh, Harvard Medical School/Mass General Brigham
Discipline
Computer Science
Start Date
17-4-2025 10:15 AM
End Date
17-4-2025 10:30 AM
Abstract
Introduction: Diagnosing fractures can be challenging in medical settings with limited expertise. Deep learning has shown promise but is restricted by image quality and. This study aims to develop a model to detect fifth metatarsal fractures using smartphone photos of radiographs. Method: A retrospective case-control study included patients >18 years with fractures (n=1240) and healthy controls (n=1224). Three-view radiographs (anterior, posterior, lateral) were obtained from the Electronic Health Record (EHR). A smartphone dataset (SP) was created using smartphone images. Deep learning models with ResNet 152V2 architecture used EHR, SP, and a combined dataset. All models were tested on an SP-test dataset. AUROC and other metrics were calculated. Continuous data were presented as median (interquartile range), with p < 0.05 considered significant. Results: The fracture group had a median age of 56 (36-68) years, while controls had 62 (51-72) years (p < 0.001). Racial composition differed significantly (Fx: 84.8% white; NoF: 92.2% white; p < 0.001). The SP model performed best (YI: 0.95, AUROC: 0.99). On SP-test data, the EHR model dropped (YI: 0.88, AUROC: 0.83), while SP and combined remained optimal. Conclusion: Deep learning models can detect fractures from smartphone photos, suggesting smartphone-trained models could aid diagnosis in resource-limited settings.
Enhancing Fracture Detection in Remote Settings: Evaluating the Efficacy of FIXUS AI Deep Learning Algorithms in Identifying Fifth Metatarsal Fractures Using Mixed-Quality X-rays
Introduction: Diagnosing fractures can be challenging in medical settings with limited expertise. Deep learning has shown promise but is restricted by image quality and. This study aims to develop a model to detect fifth metatarsal fractures using smartphone photos of radiographs. Method: A retrospective case-control study included patients >18 years with fractures (n=1240) and healthy controls (n=1224). Three-view radiographs (anterior, posterior, lateral) were obtained from the Electronic Health Record (EHR). A smartphone dataset (SP) was created using smartphone images. Deep learning models with ResNet 152V2 architecture used EHR, SP, and a combined dataset. All models were tested on an SP-test dataset. AUROC and other metrics were calculated. Continuous data were presented as median (interquartile range), with p < 0.05 considered significant. Results: The fracture group had a median age of 56 (36-68) years, while controls had 62 (51-72) years (p < 0.001). Racial composition differed significantly (Fx: 84.8% white; NoF: 92.2% white; p < 0.001). The SP model performed best (YI: 0.95, AUROC: 0.99). On SP-test data, the EHR model dropped (YI: 0.88, AUROC: 0.83), while SP and combined remained optimal. Conclusion: Deep learning models can detect fractures from smartphone photos, suggesting smartphone-trained models could aid diagnosis in resource-limited settings.