Evaluation of Deep Learning Approaches for Musculoskeletal Fracture Detection Using Radiographic Images

Abstract

Accurate and timely identification of musculoskeletal fractures in radiographs is crucial in clinical judgment particularly in emergency and orthopaedic practice. The challenges for manual interpretations of radiographic images including subtle fracture patterns, low image contrasts and inter-observer variation demonstrates the importance of automated solutions. This paper analyzes multiple Deep Convolutional Neural Networks (DCNNs) for computerized distinction of non-fractured and fractured radiographs. The images are obtained from a public dataset, the FracAtlas. DCNN models, such as ResNet50, DenseNet121, EfficientNet-B0, and EfficientNet-B3, are trained through a transfer learning strategy. A comprehensive analysis is carried out using Area Under the Curve (AUC), F1-score, precision, accuracy, and recall. The results indicate that EfficientNet-B3 obtained the maximum accuracy and F1-score of 90.7% and 71.6%, respectively. DenseNet121 achieved the highest AUC of 90.8%, whereas EfficientNet-B3 achieved an AUC of 89.6%. In general, EfficientNet-B3 and DenseNet121 performed better in classification as compared to ResNet50 and EfficientNet-B0. These results suggest that the used lightweight models can be used to facilitate automated fracture detection. Therefore, the research can be applied to implement DCNN models in practice.

Ähnliche Arbeiten