Page 50 - AIH-2-1

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

finer features across different depths. The key contributions Table 1. Reclassified pixel values for segmentation
of this study are as follows:
(1) Optimization of multi-scale skip connection paths: Vertebral bone Intervertebral disc Background regions
By refining these paths, we can better control model (100 100 100) (255 255 255) (0 0 0)
parameters while ensuring high segmentation
accuracy. biases and ensures that the training dataset captures a
(2) Increased network depth in the upsampling process: comprehensive range of variations present in the images.
This modification enables the model to reconstruct Each input spine image is standardized to a size of 512
complex structures and details more accurately, × 512 pixels to maintain consistency and optimize the
enhancing overall precision. computational efficiency during model training. Figure 1
(3) Introduction of residual connections: These shows an example of an original spinal MRI image and its
connections are incorporated locally to mitigate the corresponding label image, illustrating the visual clarity
problem of gradient vanishing, accelerate training, and distinct boundaries of the segmented regions.
and strengthen the network’s feature learning and
representation capabilities. 2.2. J-Unet architecture

In summary, the proposed method addresses the The J-Unet architecture introduced in this study represents
challenges of excessive training time and model accuracy a significant evolution from traditional U-Net and U-Net3+
in spine MRI segmentation. By optimizing the architecture designs, addressing some of the limitations inherent in
of U-Net++, we aim to provide a more efficient and these models and incorporating advanced features to
precise tool for medical professionals, thereby improving enhance performance in spine MRI image segmentation.
diagnostic and treatment outcomes for spinal diseases. Figure 2 illustrates the architecture of the proposed J-Unet,
highlighting the innovative structural elements that
2. Data and methods differentiate it from its predecessors.
In this section, the dataset used in this study and the 2.2.1. Asymmetric network architecture
processing methods employed are introduced. A detailed
description of the proposed J-Unet architecture, designed One of the most notable advancements in the J-Unet model
specifically for spine MRI image segmentation, is also is its asymmetric network architecture, which diverges
provided. from traditional U-Net structures by extending the
upsampling pathway with three additional layers compared
2.1. Dataset and processing methods to its downsampling counterpart. This asymmetry is
The dataset utilized in this study consists of T2-weighted deliberately engineered to enhance the network’s ability
MRI scans obtained from a cohort of 215 patients, sourced to capture and reconstruct complex, hierarchical features
from multiple medical institutions to ensure diversity and more comprehensively. By increasing the depth on the
robustness in the data. 28,29 All images are provided in Nifti upsampling side, the model can progressively refine the
format, a widely recognized standard for medical imaging, spatial resolution of the feature maps, thereby improving
which facilitates comprehensive 3D visualization and the accuracy of the segmentation output.
analysis. In this design, the input and output channels of
Initially, the dataset’s labels included 21 distinct pixel these additional upsampling layers are set to 120,
values representing various anatomical structures and effectively reducing the number of parameters without
features. For the purposes of this study, these original compromising accuracy. The advantages of an asymmetric
labels were reclassified into three primary categories: network architecture are evident in its enhanced ability
vertebral bone, intervertebral disc, and background to precisely reconstruct complex structures and details.
regions. This reclassification simplifies the segmentation Traditional symmetric architectures often struggle with
task, focusing on the most clinically relevant structures. images characterized by irregular shapes and intricate
The reclassified pixel values for segmentation are detailed edges. By increasing the depth on the upsampling side, the
in Table 1, providing a clear framework for subsequent asymmetric network better adapts to these complexities.
image processing and analysis. It also utilizes skip connections between specific scales
to fine-tune model parameters while maintaining
We configured the training and validation datasets
with a 4:1 ratio. To ensure a balanced and representative segmentation accuracy.
distribution of data, the allocation was performed The depth of the upsampling pathway in the J-Unet
through random sampling, which mitigates potential is crucial for reconstructing the finer details of spinal

Volume 2 Issue 1 (2025) 44 doi: 10.36922/aih.3889