Automatic Identification of Spinal Vertebrae in CT Scan Images Using A Hybrid ResNet + U-Net Neural Network

In medical diagnostics, particularly in the analysis of medical images, computational and automated methods have been employed to perform various tasks. Over the past few years, the use of artificial intelligence and machine learning-based methods has significantly improved the accuracy and effectiveness of medical diagnostics. This research presents an automated approach for identification and labeling spinal vertebrae in CT Scan images. Deep learning approaches are classified into 2D and 3D methods, each with its own strengths and computational considerations. While 3D methods often offer higher accuracy, they require more computational resources. To reduce this, creating 2D representations of 3D CT scan data has emerged as an effective strategy, narrowing the performance gap between 2D and 3D approaches. The neural network architecture presented in this study is a combination of ResNet and U-Net structures, both suitable for medical image applications. Their suitability stems from the ability to create deeper networks that mitigate the vanishing gradient problem. These structures introduce skip connections between network layers, enabling the simultaneous flow and use of information from different image dimensions. This work facilitates the use of information from all stages of downsampling. To evaluate the performance of the proposed network, the recognition rate metric was used, showing superior performance with a rate of $\mathbf{9 1 . 4 2 \%}$ compared to J-CNN, DI2IN, and BtrflyNet models in the labeling task. Notably, the integration of long and short skip connections positively impacts network performance.