Pixel-Label-Based Segmentation of Cross-Sectional Brain MRI Using Simplified SegNet Architecture-Based CNN

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of MRI images is crucial for medical diagnosis and analysis.
• Supervised learning for image segmentation often requires extensive labeled datasets, which are costly and time-consuming to acquire.

Purpose of the Study:

• To develop an efficient deep neural network approach for MRI image segmentation.
• To achieve optimal accuracy using a reduced number of training images and labels.
• To enable the application of a trained segmentation network to new, unseen MRI images.

Main Methods:

• A deep neural network model was employed for pixel-wise classification and segmentation of MRI images.
• The approach involved preprocessing MRI images and training the network on a limited dataset.
• The trained network was then applied to segment independent test images from the same database.

Main Results:

• The proposed method achieved visually high-quality segmentation results, closely resembling ground truth images.
• Quantitative evaluation showed an average Dice similarity index of approximately 0.8.
• The average Jaccard similarity measure was approximately 0.6, outperforming existing methods.

Conclusions:

• The deep neural network approach effectively segments MRI images with high accuracy.
• The method demonstrates efficiency by requiring fewer training labels, addressing the cost barrier in supervised learning.
• The validated performance indicates the potential for generating reference images highly similar to segmented ground truth.