A patch-based deep learning MRI segmentation model for improving efficiency and clinical examination of the spinal...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Magnetic resonance imaging (MRI) is crucial for diagnosing spinal diseases, including tumors.
• Conventional segmentation methods are time-consuming and inconsistent.
• Accurate segmentation is vital for diagnosis and treatment planning.

Purpose of the Study:

• To develop a fully automated segmentation method for spine MRI images.
• To enhance segmentation efficiency and accuracy for clinical applications.
• To utilize deep learning for improved spine image analysis.

Main Methods:

• A convolutional neural network (CNN) was employed for automatic feature extraction from spine MRI data.
• A patch extraction (PE) based deep neural network was developed to restore feature maps.
• Training was optimized using pre-training and an enhanced stochastic gradient descent method.

Main Results:

• The proposed method achieved high accuracy in segmenting spine MRI images, with precision of 90.6%, recall of 91.1%, and overall accuracy of 93.2%.
• The model demonstrated real-time performance, outperforming traditional segmentation algorithms.
• Key metrics included an F1-score of 91.3%, Intersection over Union (IoU) of 83.8%, and Dice Coefficient (DC) of 91.1%.

Conclusions:

• A fully automated segmentation method for spine MRI images was successfully developed using a CNN and PE-module.
• The patch extraction based neural network (PENN) effectively addresses limitations of traditional segmentation techniques.
• The method provides accurate and efficient segmentation, supporting clinical diagnosis and treatment planning for spinal conditions.