Page 53 - AIH-2-1

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

By effectively managing these complexities, the J-Unet prominent architectures in the field, thereby providing a
architecture supports more precise diagnostic outcomes clear perspective on the current state of the art in spinal
and potentially informs better clinical decision-making. MR image segmentation.
Accurate segmentation of spine MRI images can aid 3.1. Experimental setup
in the early detection and diagnosis of spinal pathologies
such as disc herniation, spinal stenosis, and tumors. Precise We utilized three evaluation metrics to assess the
delineation of these structures is essential for planning segmentation performance of our models: accuracy, mean
surgical interventions, assessing disease progression, and intersection over union (mIOU), and dice coefficient.
evaluating treatment efficacy. The improved segmentation These metrics provide a comprehensive view of the model’s
performance offered by the J-Unet model can therefore effectiveness in segmenting vertebrae and intervertebral
contribute to better patient outcomes by enabling more discs in spinal MR images. The formulas for these metrics
targeted and effective treatments. are as follows:
(1) Accuracy: Measures the proportion of true results
Moreover, the computational efficiency and scalability (both true positives [TP] and true negatives [TN])
of the J-Unet architecture make it suitable for deployment among the total number of cases examined.
in clinical settings where rapid processing of large volumes
of imaging data is required. This is particularly important Accuracy = TP +TN (I)
in modern healthcare environments, where the demand +TP +FN + FP TN
for advanced imaging techniques is increasing, and the
ability to process and analyze data quickly can significantly (2) mIOU: Calculates the average IOU across all classes,
impact the quality of care provided. providing an overall measure of segmentation
performance.
3. Results and analysis cIOU + ucIOU
In this section, we detail the organization and methodology mIOU = 2 (II)
of our experimental study using magnetic resonance (MR)
imaging data. The dataset, comprising MR image sequences (3) Dice coefficient: Evaluates the overlap between the
from 215 patients, was stratified into training and test sets predicted and ground truth segments, often used in
in a 4:1 ratio. This separation was carefully designed to medical image analysis.
ensure both sets were representative of the overall dataset,
supporting the generalizability of our findings. Using the Dice = 2TP (III)
+
J-Unet architecture, we developed a model capable of + 2TP FP FN
automatically segmenting vertebrae and intervertebral
discs within spinal MR images. The performance of this Definitions of confusion matrix terms are given below:
model was rigorously evaluated to ascertain its efficacy • TP: A positive class instance correctly predicted as
in medical imaging tasks. In addition, to provide a positive.
comprehensive analysis of our model’s capabilities, we • False negative (FN): A positive class instance
compared its performance with several other established • incorrectly predicted as negative.
False positive (FP): A negative class instance
neural network architectures: Unet, Unet++, Unet+++, incorrectly predicted as positive.
and Res-Unet. This comparative study aimed to highlight • TN: A negative class instance correctly predicted as
the strengths and potential areas for improvement in the negative.
J-Unet architecture relative to other models in handling
complex segmentation tasks in spinal MR images. In these definitions, TP and TN indicate correct
predictions of the instance class, while FN and FP
The computational experiments were conducted using
the Pytorch framework and cuDNN library, optimized for indicate incorrect predictions. Table 2 outlines the specific
parameter settings used for the segmentation model.
deep neural network operations. All models were trained
on a robust hardware setup featuring a single NVIDIA GTX These settings were carefully chosen to optimize model
3090 graphics processing unit. This high-performance performance and ensure reliable evaluation metrics. They
computing environment ensured efficient processing of
large-scale data, facilitating timely training and evaluation Table 2. Parameter settings of the segmentation model
of the models. This systematic approach allowed us to not Parameter Learning rate Optimizer Batch size Epoch
only assess the specific advantages of the J-Unet model but
also to establish a baseline for performance against other Value 1e-4 Adam 2 100

Volume 2 Issue 1 (2025) 47 doi: 10.36922/aih.3889