A novel Rician noise reduction and VGG-transformer model for lumbar spine osteoporosis detection in low radiation scans

Lumbar spine osteoporosis (LSO) is a spinal disorder caused by low bone mineral density and structural degeneration in the lower back vertebrae (lumbar spine). It leads to increased weakness and a high risk of fracture. The early detection of these fractures is vital to avoid future complications. Moreover, due to the Rician noise, mainly common in low-dose X-ray imaging, automated osteoporosis detection still poses a significant challenge for conventional diagnostic models. This research work is three-fold. Initially, we propose a novel genetic programming noise reduction algorithm to reduce the effects of Rician noise acquired from low radiation LSO X-ray images. In the second phase, we propose a novel deep hybrid model combining the VGG model for feature extraction with the transformer encoder model to capture long-range dependencies in LSO X-ray images. Finally, a fine-tuning method is utilized to customize our proposed model specifically for osteoporotic fracture detection. Our proposed model uses the VGG network to extract high-level features, local patterns, and textures related to fractures. Our model utilizes a Transformer encoder to model global linkages and dependencies in the collected features, enabling it to cover extensive areas of the image and identify small fractures. The proposed model is evaluated using publicly available LSO datasets, and the results are compared and analyzed with state-of-the-art models, achieving a F1 score of 0.92, a sensitivity of 95 %, and a specificity of 91 %. Our proposed model has exceptional reliability, as indicated by the highest Matthews Correlation Coefficient of 0.88 and Cohen's Kappa of 0.87, signifying robust concordance with the ground truth.