Abstract:

Spinal fractures are critical injuries that require prompt and accurate diagnosis to prevent long-term complications. Traditional diagnostic approaches rely heavily on manual analysis of CT images, which can be time-consuming and prone to human error. This project proposes a deep learning-based framework for automated spinal fracture detection using CT images. Two novel architectures, SmallCNN-FractureNet and SpineResNetMini, are developed to extract both local and global anatomical features, enhancing fracture detection accuracy. The dataset consists of labelled CT images of normal and fractured spines, which are preprocessed and augmented to improve model generalization. SmallCNN-FractureNet leverages a lightweight CNN to capture essential vertebral patterns, while SpineResNetMini uses residual connections to extract hierarchical features efficiently. Both models are trained with supervised learning using cross-entropy loss and optimized with Adam optimizer and cosine annealing scheduler. Evaluation metrics include accuracy, classification report, and confusion matrices, ensuring a robust assessment of model performance. The framework incorporates visualization and interpretability techniques to support clinical decision-making. This approach addresses the limitations of existing contrastive learning methods, providing an end-to-end solution for fracture classification. The system is deployed using a Flask backend with a MySQL database and a web frontend. The results demonstrate that the proposed models achieve high accuracy while remaining computationally efficient, making them suitable for integration into healthcare workflows. Overall, this project enhances diagnostic reliability and reduces the manual effort required in spinal fracture assessment, offering a scalable and interpretable solution for clinical practice.

Keywords:

Spinal Fracture, CT Images, Deep Learning, CNN, Residual Network, Feature Extraction, SmallCNN-FractureNet, SpineResNetMini, Image Classification, Medical Imaging