🤖 machine learning

Brain Tumor Classification from 3D MRI Using Persistent Homology and Betti Features: A Topological Data Analysis Approach on BraTS2020

This study proposes a topology-driven framework using persistent homology to extract interpretable Betti features from 3D FLAIR MRI volumes, achieving 89.19% accuracy in classifying high-grade and low-grade gliomas on the BraTS 2020 dataset through efficient classical machine learning classifiers.

Original authors: Faisal Ahmed

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Faisal Ahmed

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to tell the difference between two types of clouds: a fluffy, harmless cumulus cloud (a Low-Grade Glioma) and a dark, stormy, dangerous cumulonimbus cloud (a High-Grade Glioma).

Usually, doctors and computers look at these clouds by measuring how dark or light they are (intensity) or how rough their texture looks. But sometimes, two clouds can look very similar in color and texture, yet have completely different internal structures. One might be a solid block of water, while the other is full of holes and tunnels.

This paper proposes a new way to look at brain tumors using 3D MRI scans. Instead of just looking at the "color" of the tumor, the author, Faisal Ahmed, uses a mathematical tool called Topological Data Analysis (TDA). Think of this as looking at the shape of the holes inside the tumor rather than just its surface.

Here is the breakdown of the paper using simple analogies:

1. The Problem: Too Much Data, Not Enough Clarity

Brain MRI scans are like giant 3D puzzles made of millions of tiny cubes (voxels). Traditional AI (Deep Learning) tries to solve this puzzle by memorizing millions of examples. It's like trying to learn a language by memorizing every single sentence in a library. It works, but it requires a massive library (lots of data), a huge brain (supercomputers), and the final result is often a "black box"—we know it works, but we don't know why.

2. The Solution: The "Hole" Detector

The author suggests a smarter, lighter approach. Instead of memorizing every pixel, let's count the holes.

In mathematics, there are three main types of "holes" or shapes we can find in a 3D object:

Betti-0 (The Islands): How many separate pieces are there? (Is the tumor one big blob, or is it broken into three tiny islands?)
Betti-1 (The Loops): Are there any tunnels or rings? (Like a donut or a tunnel through a mountain.)
Betti-2 (The Caves): Are there any enclosed empty spaces inside? (Like a hollow ball or a cave inside a rock.)

The author uses a method called Persistent Homology. Imagine slowly filling the MRI scan with water, starting from the bottom up. As the water rises:

New islands appear (Betti-0).
Tunnels form and then get flooded (Betti-1).
Caves get filled up (Betti-2).

By tracking when these shapes appear and when they disappear, the computer creates a unique "fingerprint" of the tumor's internal structure.

3. The Process: From 3D Scans to a Simple List

Here is what the author actually did:

The Data: They used 3D MRI scans of brains from the BraTS 2020 dataset (a famous collection of brain tumor scans). They focused on the FLAIR scan, which is like a special filter that makes the tumor "glow" against the healthy brain.
The Filter: They didn't look at the whole brain, just the middle slice where the tumor is usually found, to save time.
The Counting: They ran the "water filling" simulation (Persistent Homology) to count the islands, loops, and caves at 100 different levels of "water height."
The Result: This turned a massive 3D image into a simple list of 300 numbers. These numbers describe the tumor's shape perfectly.
The Cleanup: Some of those 300 numbers were just noise. The author used a "feature selection" tool (like a sieve) to keep only the top 14 to 42 most important numbers that actually helped tell the difference between the two tumor types.

4. The Test: Simple Math vs. Complex AI

The author took these simplified lists of numbers and fed them into two classic, easy-to-understand machine learning models: Random Forest (a committee of decision trees) and XGBoost (a smart, step-by-step learner).

The Results:

The system didn't need a supercomputer. It ran on a standard laptop.
It didn't need to "fake" extra data (data augmentation) to learn.
Accuracy: The best model (Random Forest) got it right 89.19% of the time.
Interpretability: Because the model is based on counting holes and loops, a doctor can actually understand why the computer made a decision. "It's a High-Grade tumor because it has too many internal caves," is a much clearer explanation than "The neural network's hidden layer 4 activation suggests a tumor."

The Big Takeaway

Think of this paper as swapping a massive, expensive, black-box camera for a smart, lightweight pair of glasses.

Instead of trying to memorize the entire picture of the brain, this method teaches the computer to look for the shape of the holes. It proves that you don't always need a giant, complex AI to solve medical problems. Sometimes, a clever mathematical way of looking at the "shape" of the data is faster, cheaper, and easier to trust.

In short: The author showed that by counting the "islands, tunnels, and caves" inside a brain tumor, we can accurately tell if it's dangerous or not, using a simple computer and a clear, logical explanation.

1. Problem Statement

Accurate classification of brain tumors (specifically distinguishing between High-Grade Glioma [HGG] and Low-Grade Glioma [LGG]) from 3D Magnetic Resonance Imaging (MRI) is challenging due to:

High Dimensionality: 3D MRI volumes contain vast amounts of data, making processing computationally expensive.
Complex Structural Patterns: Tumors exhibit intricate morphological variations that are difficult to capture with standard intensity-based features.
Limitations of Deep Learning: While Convolutional Neural Networks (CNNs) are dominant, they require massive labeled datasets, extensive data augmentation, and significant computational resources. Furthermore, they often lack interpretability, which is critical for clinical decision-making.

The paper proposes a solution that bypasses these limitations by using Topological Data Analysis (TDA) to extract compact, interpretable, and robust structural descriptors directly from 3D MRI volumes.

2. Methodology

The proposed framework is a "topology-driven" pipeline that converts 3D MRI volumes into fixed-length feature vectors using Persistent Homology. The process consists of four main stages:

A. Data Preprocessing

Dataset: BraTS 2020 dataset containing 369 subjects (293 HGG, 76 LGG).
Modality: Fluid Attenuated Inversion Recovery (FLAIR) images were selected as they best highlight tumor-associated edema.
Volume Reduction: To reduce computational load while preserving tumor structures, the central 60 slices (indices 30–90) of the original 155-slice volumes were extracted, resulting in a working volume of approximately $60 \times 240 \times 240$ .

B. Filtration Construction

A cubical sublevel filtration is constructed from the grayscale MRI volume.
The intensity range $[0, 255]$ is discretized into $N=100$ thresholds ( $t_1 < t_2 < \dots < t_{100}$ ).
As the threshold increases, voxels are progressively activated, creating a nested sequence of binary images ( $X_1 \subset X_2 \subset \dots \subset X_{100}$ ). This tracks how topological structures emerge and evolve.

C. Persistent Homology & Feature Extraction

Persistent Homology tracks the "birth" and "death" of topological features across the filtration.
Betti Numbers: The method computes Betti numbers for three homology dimensions:
- Betti-0 ( $k=0$ ): Connected components (number of distinct regions).
- Betti-1 ( $k=1$ ): Loops or tunnels.
- Betti-2 ( $k=2$ ): Voids or cavities.
Vectorization: Instead of using raw persistence diagrams (which are sets of points), the authors convert them into Betti curves. For each dimension $k$ , a vector $\beta_k(X)$ of length 100 is generated, representing the count of features at each threshold.
Final Feature Vector: The vectors for $k=0, 1, 2$ are concatenated to form a 300-dimensional feature vector ( $f(X) \in \mathbb{R}^{300}$ ) for each MRI volume.

D. Classification

Feature Selection: To reduce redundancy, feature importance scores from tree-based models are used to select a subset of the most informative Betti features.
Classifiers: Two classical machine learning models were trained:
1. Random Forest (RF)
2. Extreme Gradient Boosting (XGBoost)
Task: Binary classification of HGG vs. LGG.
Advantage: The topological features are inherently invariant to rotation, reflection, and minor intensity variations, eliminating the need for data augmentation.

3. Key Contributions

Topology-Driven Framework: Proposes a novel approach applying persistent homology directly to full 3D MRI volumes (rather than 2D slices) for tumor classification.
Interpretable Descriptors: Extracts Betti-based features (0, 1, 2) that provide a mathematical description of tumor morphology (components, loops, voids), offering high interpretability compared to "black box" deep learning features.
Computational Efficiency: Demonstrates that high performance can be achieved without complex deep learning architectures, massive datasets, or data augmentation.
Dimensionality Reduction: Successfully reduces 3D volumetric data to a compact 300-dimensional (or fewer with selection) feature set while preserving critical structural information.

4. Experimental Results

Experiments were conducted on the BraTS 2020 dataset with an 80/20 train-test split.

Feature Selection Impact: Feature selection significantly improved performance and reduced the number of features required.
- XGBoost: Achieved 87.84% accuracy using 42 selected features (combined B0+B1+B2).
- Random Forest: Achieved the best overall performance of 89.19% accuracy.
  - The best RF configuration used a combination of B1+B2 features (42 features) or all three (14 features).
  - Precision reached 90.50% for the B1+B2 combination.
Comparison:
- Combining all Betti features (B0+B1+B2) generally outperformed using single features.
- The approach achieved competitive results against deep learning methods while using a fraction of the computational resources and requiring no data augmentation.
Visualization: PCA plots and Betti curves clearly showed separable distributions between HGG and LGG, confirming that topological descriptors capture distinct structural differences between tumor grades.

5. Significance and Conclusion

Clinical Relevance: The method offers a lightweight, interpretable alternative to deep learning, which is crucial for medical settings where understanding why a model makes a decision is as important as the decision itself.
Robustness: The topological approach is robust to noise and geometric transformations, making it suitable for real-world clinical data where image quality may vary.
Efficiency: By avoiding the need for large-scale training data and heavy GPU resources, this framework is more accessible for deployment in resource-constrained environments.
Future Work: The authors suggest integrating these topological descriptors with deep learning representations and extending the framework to multi-modal MRI data for even richer analysis.

In summary, the paper successfully demonstrates that Persistent Homology can effectively distill complex 3D brain tumor structures into simple, interpretable numerical features, enabling highly accurate classification with standard machine learning algorithms.