Automated glioblastoma segmentation based on a multiparametric structured unsupervised classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Supervised learning methods for brain tumor segmentation require extensive labeled data, which is time-consuming to acquire.
• Unsupervised methods offer an alternative but often achieve lower accuracy compared to supervised approaches.

Purpose of the Study:

• To develop an automated unsupervised method for brain tumor segmentation using anatomical Magnetic Resonance (MR) images.
• To evaluate the performance of different unsupervised classification algorithms for this task.

Main Methods:

• The study evaluated four unsupervised algorithms: K-means, Fuzzy K-means, Gaussian Mixture Model (GMM), and Gaussian Hidden Markov Random Field (GHMRF).
• An automated post-processing step using statistical tissue probability maps was employed to identify tumor classes.
• The methods were validated using the public BRATS 2013 Test and Leaderboard datasets.

Main Results:

• The Gaussian Mixture Model (GMM) approach achieved results comparable to and exceeding most supervised methods on the Leaderboard set, securing second place.
• The Gaussian Hidden Markov Random Field (GHMRF) variant ranked first among unsupervised methods and seventh overall on the Test set.
• These results demonstrate the effectiveness of the proposed unsupervised segmentation pipeline.

Conclusions:

• The proposed automated unsupervised method provides a viable and effective alternative for brain tumor segmentation.
• Unsupervised approaches, particularly GMM and GHMRF, can achieve high performance without the need for labeled training data.