Advanced Methods for Addressing Imbalanced Datasets in Bone Fracture Detection

Managing imbalanced datasets remains a key problem in machine learning, especially in areas such as healthcare, fraud detection, and industrial applications. Traditional methods such as random oversampling, undersampling, and data augmentation often fail to effectively balance datasets, often introducing redundancy and noise. These shortcomings result in models that are biased and generalize poorly. In this article, we present a dual-strategy approach that is tailored to the degree of class imbalance. For datasets with mild imbalance, the proposed method involves DOWNSAMPLING and UPWEIGHTING the majority class, ensuring a simpler yet effective solution. For extreme imbalance, a more robust combination of Adaptive Synthetic sampling (ADASYN) and Neighborhood Cleaning Rule (NCL) is used. ADASYN generates synthetic samples of the minority class, while NCL removes noisy samples from the majority class, thus improving the quality of the dataset. The proposed method is evaluated on a fracture dataset that exhibits severe class imbalance, with intact bone data accounting for only 5% of the total data. Results show substantial improvements across multiple metrics, including Recall, F1 Score and the AUC Curve. These findings demonstrate the efficacy of the proposed approach in addressing both mild and extreme class imbalance.