Deep Learning-based Binary Classification of Bone Fractures using a Hybrid MobileNetV3-CNN Architecture and Clinical X-ray Dataset

Abstract

Identifying the presence of bone fractures from radiographic images is a clinically important and time-consuming task that is limited by inter-observer variability or availability of radiologists. This research proposes an end-to-end deep learning framework for the automated classification of fractures using a hybrid architecture of MobileNetV3-Small and specifically designed new convolutional layers. The framework was trained and validated on the Fracture Multi-Region X-ray dataset, a publicly available dataset containing 10,580 grayscale X-ray images labelled from different anatomically connected regions (bones of limbs, spine, hips, knees). To improve diagnostic accuracy and freedoms, we implemented transfer learning, data preprocessing, a comprehensive and internal augmentation pipeline (including flipping, rotation, contrast, and zoom), etc. MobileNetV3-Small was used as a lightweight feature extractor and further optimized with our convolutional blocks, batch normalization, dropout, and global average pooling. Our training strategy included two separate phases, where the first phase aimed to solely extract the features, with the second phase featuring feature extraction and then fine-tuning, while optimization was executed using AdamW with a fully adaptive learning schedule. Evaluation performance on the test set resulted in isolated classification accuracy of 97%, and was candidates set against DenseNet121, VGG-16, and ResNet50. The findings validate the effectiveness of using lightweight architectures enhanced with domain-specific features for precise medical image analysis. Due to its computational efficiency, the model is suitable for real-time and resource-constrained applications, such as point-of-care and mobile health systems. This work establishes a foundation for future improvement through transformer-based architectures, multimodal data types, and explainable AI to facilitate diagnostic confidence and opportunities for clinical decision-making.

Ähnliche Arbeiten