SliceInsightNet: Advanced Hybrid Image Segmentation for Spine Annotation in MRI

Segmentation in medical imaging is crucial for identifying and delineating anatomical structures, detecting pathological regions, and planning treatments. Deep learning techniques is the current state-of-the-art in several medical imaging segmentation applications. Three-dimensional (3D) data processing is known to efficient in handling changes in axial direction, however, it is computationally expensive. Processing images as a set of two-dimensional (2D) slices is more computationally efficient but with the cost of losing inter-slicing knowledge. This study focuses on the challenging task of spine segmentation, essential for diagnosing vertebral abnormalities and planning surgical interventions. We present SlicelnsightNet, a novel hybrid approach that leverages both 2D and 3D segmentation techniques by utilizing sliced image data, optimizing computational efficiency and accuracy. Kullback-Leibler (KL) divergence is employed to calculate similarity metrics. The proposed model demonstrated slightly better results on unseen window configurations compared to UNet, and consistently outperformed UNet++ in both seen and unseen cases. Specifically, SlicelnsightNet achieved a Dice Coefficient (DC) of 0.9712, Intersection over Union (IoU) of 0.9441, Pixel Accuracy of 0.9951, and F1 Boundary of 0.9848 for the same window configuration using the open access lumber spine segmentation dataset. For different window configurations, SlicelnsightNet maintained competitive performance with a DC of 0.9171, IoU of 0.8475, Pixel Accuracy of 0.9881, and F1 Boundary of 0.9666. These results confirm SlicelnsightNet's superior performance in spine segmentation, reinforcing its potential for broader clinical applications.