A Novel Ensemble Approach for Rib Fracture Detection and Visualization using CNNs and Grad-CAM.

Abstract

Aim: This study aimed to develop a reliable and efficient system for predicting and locating rib fractures in medical images using an ensemble of convolutional neural networks (CNNs).

Methods: We employed five CNN architectures-Visual Geometry Group Network 16 (VGG16), Densely Connected Convolutional Network 169 (DenseNet169), Inception Version 4 (Inception V4), Efficient Network B7 (EfficientNet-B7), and Residual Network Next 50 layers (ResNeXt-50)-trained on a dataset of 840 grayscale computed tomography (CT) scan images in .jpg format collected from 42 patients at a local hospital. The images were categorized into two groups representing healed and fresh fractures. The ensemble model was designed to improve predictive accuracy and robustness, utilizing techniques like gradient-weighted class activation mapping (Grad-CAM) for visualization of fracture locations.

Results: The ensemble model achieved an accuracy of 0.96, area under the curve (AUC) of 0.97, recall of 0.97, and F1 score of 0.96. Grad-CAM visualizations could effectively locate rib fractures, providing crucial assistance in diagnostics.

Conclusions: The ensemble model demonstrates high accuracy and robustness in fracture detection, underscoring its potential for enhancing diagnostic processes in clinical settings. Despite limitations such as the small dataset size and lack of diverse demographic representation, the results are promising for future clinical application.