WristNet: A Novel Deep Learning Approach for Detecting Fractures in Wrist X-ray Images.

Wrist fractures are some of the most frequent injuries in emergency medicine, and correct interpretation is demanded for clinical management. Recent advances in deep learning offer promising solutions. However, their performance is typically limited by the lack of medical data and is architecture type-sensitive. In this work, we propose WristNet, a novel deep learning approach that fuses DenseNet169 and InceptionV3 through feature-level concatenation for classifying wrist fractures. The proposed approach leverages DenseNet's ability to preserve fine structural details and InceptionV3's multiscale feature extraction to construct a rich and discriminative representation. It was evaluated on an extended version of a publicly available wrist X-ray dataset with data augmentation. It is compared with five state-of-the-art convolutional neural networks, showing a higher accuracy of 99.63%. In addition, the study shows that data augmentation does not help all architectures equally, emphasizing the need for architectureaware training strategies. Our approach exhibits strong potential to be incorporated into a computer-aided diagnostic system for ensuring the detection of wrist fractures.