Page 51 - AIH-2-1

Artificial Intelligence in Health Asymmetric U-Net for enhanced spinal MRI segmentation

Figure 1. Example original spinal magnetic resonance imaging image and its corresponding label image

Figure 2. The architecture of the proposed J-Unet
structures, which are often lost in traditional symmetric to leverage multi-scale feature information effectively.
architectures. This extended pathway allows the network Fully-scaled skip connections can introduce excessive
to learn more detailed and nuanced representations of redundancy, leading to an unnecessary increase in
the spinal structures, crucial for tasks that require high computational load and model complexity. In contrast,
precision, such as medical image segmentation. The ability adjacent-scale skip connections streamline the network,
to capture and reconstruct hierarchical features ensures reducing computational overhead without sacrificing the
that the segmentation of the spine is both accurate and richness of the multi-scale features. This optimization not
reliable, which is essential for clinical applications where only simplifies the network architecture but also enhances
precise anatomical delineation is required for diagnosis computational efficiency and model scalability, making the
and treatment planning. J-Unet more practical for large-scale applications and real-
time processing.
2.2.2. Adjacent-scale skip connections
In the J-Unet model, the size of feature maps does
Unlike the U-Net3+ architecture, which utilizes fully- not uniformly change by integer multiples. For instance,
scaled skip connections, the J-Unet adopts adjacent-scale adaptive max pooling is employed to resize a 256 × 256
skip connections. This approach strategically reduces feature map from the encoder to a 192 × 192 feature map in
the redundancy and overall parameter count associated the decoder, whereas standard fixed window max pooling
with the skip connections while maintaining the ability is used for regular size adjustments. This method enhances

Volume 2 Issue 1 (2025) 45 doi: 10.36922/aih.3889