Convolutional Neural Networks Accurately Diagnose Lisfranc Injuries on Weightbearing and Nonweightbearing Radiographs

Category: Other; Midfoot/Forefoot Introduction/Purpose: Convolutional neural networks (CNNs) predict Lisfranc injuries on weightbearing radiographs. Lisfranc injuries can be misdiagnosed 20%-50% of the time, thus leading to long-term sequela such as midfoot arthritis. Often Lisfranc injuries present to the Emergency department or general practitioner's office where Lisfranc injuries are more frequently misdiagnosed, and weightbearing films are less frequently obtained. CNNs are utilized to classify images, in particular, CNNs are successful at radiographic diagnosis of fractures, Lisfranc injuries, and classification of medical imaging. We aim to utilize CNNs to accurately diagnose Lisfranc injuries on both weightbearing and nonweightbearing films in a generalizable population of patients presenting to the emergency department, general practitioner, and orthopaedic clinics. Hypothesis: CNNs will accurately predict Lisfranc Injuries on weightbearing and nonweightbearing radiographs. Methods: In a retrospective study of 120 patients, 73 with Lisfranc injuries, and 47 without Lisfranc injuries were identified through medical record and imaging review. Radiographs and chart review were utilized to diagnose patients into Lisfranc and Not Lisfranc cohorts. A resnet-50 CNN model was trained, validated, and tested in a ratio (70:20:10) for each imaging view (anteriorposterior, lateral, oblique). Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, F score, and area under the curve (AUC) were calculated for each model and compared between weightbearing and nonweightbearing films. Results: The best deep learning model for Lisfranc identification has 94.5% sensitivity, 78.7% specificity, 90.2% negative predictive value, 87.3% positive predictive value, 88.3% accuracy, 90.8% F score, and 91.5% AUC. Only 2 out of 48 Lisfranc injuries were missed on nonweightbearing films and 3 out of 25 Lisfranc injuries were missed on weightbearing films. Conclusion: CNNs identify Lisfranc injuries on weightbearing and nonweightbearing radiographs with acceptable accuracy. The best model missed fewer Lisfranc injuries than frequently reported numbers for trained radiologists. F score and AUC indicate that this is an excellent test. This provides a method to improve and supplement diagnosis in patients with acute foot injuries, thus limiting long-term sequela from misdiagnosed Lisfranc injuries.