CNN-based radiographic acute tibial fracture detection in the setting of open growth plates

Abstract

Pediatric tibial fractures are commonly diagnosed by radiographs and constitute one of the common tasks performed by pediatric radiologists. Here, we assess the performance of a convolutional neural network for the detection of acute tibial fractures trained with a limited number of cases in skeletally immature patients. This retrospective study was performed on radiology reports manually classified as normal or tibial fracture. Classified images of orthopaedic implants, casting, and images including other pathology were excluded. The remaining cases constituted 516 studies containing 2118 radiographs. These radiographs were truncated to include a limited investigated field of view which included the distal third of the leg, inclusive of the distal physis. After exclusions, the culled dataset was randomly divided into a training set containing 784 radiographs, a validation set containing 98 radiographs, and a test set 98 radiographs. We used a modified transfer learning approach based on the Xception architecture with additional fully convoluted reasoning and drop-out layers. Of 49 fractures, two were misdiagnosed as normal. Of 49 normal exams, none were misdiagnosed. This led to model accuracy of 97.9%, sensitivity 95.9%, and specificity 100%, comparable to or better than human radiologists. In no instances were normal physes or normal developmental epiphyseal fragmentation of the tibial tuberosity or medial malleolus misclassified as a fracture. We report an efficient method to use a pre-trained network and adapt it to a medical classification task using only a small number of radiographs dedicated to precise anatomical location.