Grey Wolf Optimization Guided Non-Local Means Denoising for Localizing and Extracting Bone Regions from X-Ray Images

Abstract

The key focus of the current study is implementation of an automated semantic segmentation model to localize and extract bone regions from digital X-ray images. Methods: The proposed segmentation framework uses a pre-processing stage which follows convolutional neural network (CNN) obtained segmentation stage to extract the bone region from X-ray images, mainly for diagnosing critical conditions such as osteoporosis. Since the presence of noise is critical in image analysis, the X-ray images are initially processed with a grey wolf optimization (GWO) guided non-local means (NLM) denoising. The segmentation stage uses a Multi-Res U-Net architecture with attention modules. Findings: The proposed methodology shows superior results while segmenting bone regions from real X-ray images. The experiments include an ablation study that substantiates the need for the proposed denoising approach. Several standard segmentation benchmarks such as precision, recall, Dice-score, specificity, Intersection over Union (IOU), and total accuracy have been used for a comprehensive study. The proposed architectural has good impact compared to the state-of-the-art bone segmentation models and is compared both quantitatively and qualitatively. Novelty: The denoising using GWO-NLM adaptively chose the denoising parameters based on the required conditions and can be reused in other medical image analysis domains with minimal finetuning. The design of the proposed CNN model also aims at better performance on the target datasets.