Image Pre-processing for Deep Learning in Bone Fracture Classification: A Comparative Study

Abstract

Accurate and timely diagnosis of bone fractures is paramount for effective clinical decision-making, and deep learning models offer a promising avenue for automating this task with high precision. This study provides a comprehensive comparative analysis of the under-explored role of image pre-processing on the performance of Convolutional Neural Network (CNN) models for bone fracture classification. We evaluated a suite of architectures, including ResNet50, VGG16, EfficientNetB0, and various two-and three-model ensembles, on a bone fracture dataset. The images were subjected to diverse pre-processing pipelines, from no modification to complex combinations involving segmentation, Contrast Limited Adaptive Histogram Equalization (CLAHE), denoising, and edge detection. Our findings reveal that a tailored pre-processing pipeline consisting of raw images, CLAHE, and Gaussian blur consistently achieved the highest accuracy of 89% on DenseNet121 architecture and accuracy of 88% on Xception, MobileNetV2 model. Notably, this study also highlights the inherent robustness of modern architectures, as ResNet50 achieved a commendable 88% accuracy with only minimal image processing. Using comprehensive metrics including precision, loss, confusion matrices, ROC curves, and Grad-CAM visualizations, we demonstrate that while targeted pre-processing enhances performance, overly aggressive filtering can diminish accuracy by obscuring subtle yet diagnostically crucial fracture features. This research underscores the nuanced importance of pre-processing and reinforces that optimal results in fracture detection are achieved by balancing advanced CNN architectures with carefully selected image enhancement techniques.

Ähnliche Arbeiten