Enhancing U-Net for Wrist Fracture Segmentation in X-ray Images using Adaptive Callbacks and Weighted Loss Functions

The detection of wrist fracture through medical imaging is causing considerable challenges due to the subtle and variable manifestation of such ruptures, necessitating precise and reliable segmentation methods. Therefore, this research aimed to propose an improved U-Net model for detecting wrist fracture. The model incorporated two innovations, namely adaptive callback training and weighted loss combination. The adaptive callback mechanism could be performed by dynamically adjusting the training parameters based on the model performance to prevent overfitting and accelerate convergence. At the same time, the loss function combined Dice Loss and Binary Cross-Entropy (BCE) Loss with linear as well as non-linear exponential weighting strategies, ensuring balanced optimization between region-based accuracy and pixel classification. During this analysis, a series of experiments were conducted on a curated wrist X-ray image dataset, and the results showed that the proposed method expressed superior performance in terms of segmentation accuracy when compared with previous U-Net and other state-of-the-art procedures. The proposed method achieved 91% accuracy, 87% precision, 86% recall, and 87% F1 score. Following this discussion, the findings showed the efficacy of the adaptive training design and loss function in improving the strength and sensitivity of the model in detecting wrist fracture