Automated Detection of Bone Fractures in Muscle X-Ray Images Using Multiband-Frequency Aware Deep Representation Learning

The automated detection of bone fractures in muscle X-ray (MXR) images using artificial intelligence is vital for successful treatment and better patient outcomes. This paper proposes the multiband-frequency aware deep representation learning network (MFADRLN)-based automated approach for detecting bone fractures in MXR images. The discrete wavelet transform-based multiresolution analysis is utilised to evaluate subband images from the MXR image. Then, the deep representation learning (DRL) is applied to each subband of the MXR image, followed by feature concatenation, a dense layer, and a sigmoid layer for detecting bone fractures. The DRL branch for each subband mainly consists of the pre-trained or frozen EfficientNetV2B2-based block, a flattened layer, two successive dense-batch normalisation (BN)-dropout layer blocks, followed by the sigmoid layer for extracting multi-band frequency-aware features from MXR images. The MFADRLN model's importance is to capture the frequency-specific and spatial information of the MXR image and obtain an improved feature representation for efficient detection of bone fractures. The publicly available musculoskeletal X-ray image databases are used to evaluate the performance of the proposed MFADRLN-based approach. The results reveal that the MFADRLN has obtained the accuracy and F1-score values of 92.22% and 0.841, respectively, for detecting bone fractures. The proposed approach has demonstrated superior performance compared to the existing transfer learning techniques (ResNet50, EfficientNetV2B2, DenseNet201, MobileNetV2, InceptionV3, and XceptionNet), Vision transformer and swin transformer models to detect bone fractures in MXR images from the same database. The classification performance of the MFADRLN is compared with existing deep-learning techniques for detecting bone fractures in MXR images.