A NEW AUGMENTATION TECHNIQUE FOR IMPROVED BONE FRACTURES IMAGE CLASSIFICATION USING YOLO

Abstract

The classification of bone fractures in medical imaging distinguishing between fractured and non-fractured bones is one of the most critical and difficult tasks. Traditional manual examination is prone to errors, especially in cases of tiny or minute fractures. Similarly, these tiny fractures are a big challenge for deep learning models, which struggle to detect and classify them effectively due to their size and variability. In this paper, we propose a new augmentation method called the Cropping Small Fracture Patches (CSFP) aimed at enhancing the representation of small fractures within training data. Our approach integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance image contrast, followed by a targeted patch-based augmentation strategy that extracts small fracture regions, upscales them using Lanczos interpolation, and then combines them with original train images to enrich the dataset and used to train the YOLOv8m classification model. The proposed methodology was applied to the FracAtlas dataset and the result demonstrates a significant improvement in classification performance achieving 100% accuracy outperforming previous methods. These findings confirm the effectiveness of this approach in improving model performance; by enhancing small fracture visibility, reducing class imbalance, and improving the diversity of the dataset. However, certain limitations exist. The augmentation method relies on bounding box annotations, which may not always be available in classification datasets. Additionally, this study currently focuses on binary classification (fractured vs. non-fractured) and does not classify fracture types or anatomical locations

Ähnliche Arbeiten