Utilizing Deep Convolutional Generative Adversarial Networks for Automatic Segmentation of Gliomas: An Artificial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain MRI analysis is crucial for diagnosing neurological conditions.
• Automated methods are needed to improve efficiency and accuracy in detecting brain abnormalities.
• Deep convolutional generative adversarial networks (DCGANs) offer potential for image analysis tasks.

Purpose of the Study:

• To develop and evaluate a DCGAN model for detecting brain MRI anomalies, specifically gliomas.
• To enable simultaneous segmentation and anomaly detection using a single model.
• To assess the model's performance when trained exclusively on normal brain MRI data.

Main Methods:

• A DCGAN model was trained using T2-weighted MRIs from 300 healthy subjects.
• The model learned to generate normal synthetic brain MRI images.
• Anomaly detection was performed by comparing test MRIs against generated normal images, calculating residual loss for segmentation and abnormality identification.

Main Results:

• The model achieved high accuracy (0.907), precision (0.892), recall (0.923), and AUC (0.907).
• It correctly identified anomalies in 24 out of 27 high-grade glioma (HGG) patient MRIs.
• The model correctly classified 25 out of 27 healthy subjects' MRIs as normal.

Conclusions:

• DCGAN models trained solely on normal brain MRIs can effectively detect abnormalities.
• The proposed model demonstrates a unique approach by learning only normal patterns to identify deviations.
• This method shows promise for automated, accurate detection of brain tumors and other anomalies in MRI scans.