A Three-Phase Explainable Deep Learning Approach for Reliable Wrist Fracture Identification from X-Ray Images

Abstract

Wrist fractures present significant challenges in clinical diagnosis, often leading to treatment delays and compromised recoveries. Manual diagnosis is resource-intensive and error-prone. To address these challenges, we develop DeepWristFNet, a compact convolutional architecture designed for end-to-end wrist fracture classification using a small dataset of 193 wrist X-ray images. The DeepWristFNet architecture integrates multi-scale convolutional operations with Fire and Shuffle modules within a compact network design, followed by fully connected layers for binary classification. We applied data pre-processing techniques such as data augmentation, image enhancement, and image resizing to increase the number of images, improve image quality, and resize images to match the DeepWristFNet input size. The proposed method comprised three phases. In the first phase, we trained, validated, and tested end-to-end and achieved validation and testing accuracies of 99.04% and 87.93%, respectively. Testing was performed on a hold-out subset of image instances that was kept separate from model development. The evaluated hold-out images originated from the same dataset distribution and included the corresponding augmented variants. In the second phase, we further evaluated the learned representation by extracting deep features from the first fully connected layer of DeepWristFNet. ReliefF was then used to select informative features, which were subsequently evaluated using 10 conventional machine learning classifiers. Out of 10 classifiers, 5 classifiers, i.e., Efficient linear SVM, quadratic SVM, Narrow NN, wide NN, and medium NN, achieved 100% testing accuracy on unseen samples. In the third phase, an auxiliary Fuzzy Inference System provides an intensity-based foreground-background representation of the X-ray images. This representation provides complementary visual information for interpretation but is not intended to directly classify or localize fractures. Grad-CAM is additionally used to visualize image regions contributing to the DeepWristFNet predictions, providing a model-specific explanation of the classification decision. Additionally, we evaluated how well the proposed DeepWristFNet approach performed against cutting-edge deep transfer learning models. In the evaluated experiments, DeepWristFNet outperformed the compared pre-trained deep learning architectures on the unseen hold-out subset from the same dataset distribution (test set). This study demonstrates the potential of DeepWristFNet for wrist fracture classification under a small-data setting. However, further evaluation on larger, independently collected clinical datasets is required to establish its robustness, generalizability, and suitability for clinical decision support.