Developing a Deep Learning Model Using Transfer Learning from EfficientNet-b3 to Detect Knee Fracture on X-ray Images

Conventional radiographs are used for fracture detection routinely in knee injury patients. Miss diagnosis is harmful to patients and stressful to physicians. Thus, a clinical decision support system utilizing a deep neural network should be helpful in preventing physicians from overlooking and also improving patient safety. This study uses a deep learning model (DLM) with transfer learning from EfficientNet-b3 to detect knee fractures on X-ray images. About 12% of the total 13,615 cases were used to test the model. The testing accuracy of the trained model was 90.56%. The area under the receiver operator characteristic curve (AUC) was 0.960. Our findings highlight that the deep learning model can detect knee fractures with remarkable performance. Further implementation into clinical use as a decision support system can be helpful to prevent misdiagnosis and subsequent patient harm.