🤖 machine learning

Topological Signatures vs. Gradient Histograms: A Comparative Study for Medical Image Classification

This study compares Histogram of Oriented Gradients (HOG) and Topological Data Analysis (TDA) for retinal fundus image classification, finding that while both feature extraction methods yield similar performance using XGBoost, they provide complementary structural representations for detecting and grading diabetic retinopathy.

Original authors: Faisal Ahmed

Published 2026-02-10

📖 3 min read☕ Coffee break read

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 a detective trying to identify a specific type of counterfeit coin. To catch the fake, you could use two completely different methods: one method looks at the tiny scratches and textures on the surface, while the other looks at the overall shape and whether there are any strange holes or structural flaws in the metal.

This research paper is essentially a "detective's manual" comparing two different ways of looking at medical images (specifically, photos of the back of the eye) to detect Diabetic Retinopathy, a disease that can cause blindness.

The researcher, Faisal Ahmed, compared two "detective tools": HOG and TDA.

1. The Two Detectives

Detective HOG (The Texture Specialist):
Think of HOG like a magnifying glass that looks for patterns and edges. If you were looking at a forest, HOG wouldn't care about the shape of the forest; it would focus on the direction the blades of grass are leaning or the sharp edges of the tree trunks. In medical terms, HOG looks at the tiny textures, lines, and "grain" of the eye to see if something looks "off."

Detective TDA (The Shape Specialist):
Think of TDA like a sculptor who cares about structure and connectivity. Instead of looking at tiny scratches, TDA asks: "Is this shape solid? Does it have weird holes in it? Are there disconnected islands of color where there shouldn't be?" It looks at the "topology"—the big-picture skeleton of the image. It’s less interested in the texture of the skin and more interested in whether the "map" of the eye has strange, unexpected loops or gaps.

2. The Experiment

The researcher took a massive collection of eye photos (the APTOS dataset) and gave them to these two detectives. He then used seven different "AI assistants" (machine learning models) to see which detective provided the best clues for:

Task A: Is the eye healthy or diseased? (Yes/No)
Task B: How bad is the disease? (Level 1, 2, 3, 4, or 5)

3. The Results: Who Won?

HOG (The Texture Specialist) was the star performer. It was generally more accurate, especially when it came to the "Yes/No" question. Because it captures so much tiny, granular detail (it actually creates over 26,000 tiny clues per image!), it was very good at spotting the subtle textures of the disease.
TDA (The Shape Specialist) was the "smart" alternative. While it didn't quite beat HOG in raw accuracy, it was much more "interpretable."

Why does "interpretable" matter?
Imagine an AI tells a doctor, "This patient is sick," but can't explain why. That's scary.

HOG is like a detective saying, "I see a pattern of 5,000 tiny scratches that look suspicious." (Hard for a human to visualize).
TDA is like a detective saying, "I see that there are twice as many 'holes' or bright spots in this image compared to a healthy eye." (Very easy for a doctor to understand and verify).

4. The Big Picture (The "Hybrid" Future)

The paper concludes that we shouldn't have to choose between them.

The best way forward is a "Super-Detective" approach. By combining the high-detail texture clues from HOG with the structural, easy-to-understand shape clues from TDA, we can create AI tools for doctors that are both incredibly accurate and easy to trust. It’s like having a detective who can see both the microscopic scratches on a coin and the overall shape of the metal at the same time.

Technical Summary: Topological Signatures vs. Gradient Histograms

Title: Topological Signatures vs. Gradient Histograms: A Comparative Study for Medical Image Classification
Author: Faisal Ahmed
Focus: Comparative analysis of handcrafted feature extraction for Diabetic Retinopathy (DR) detection.

1. Problem Statement

While deep learning has revolutionized medical imaging, many models suffer from a lack of interpretability and generalizability. Handcrafted features are often used to bridge this gap, but there is a lack of systematic research comparing different types of descriptors. Specifically, no prior study has directly compared gradient-based descriptors (which capture local texture/edges) with Topological Data Analysis (TDA) (which captures global shape/connectivity) on the same medical imaging dataset to determine their individual and complementary strengths.

2. Methodology

The study utilizes the large-scale APTOS 2019 dataset (retinal fundus images) to perform two distinct classification tasks:

Binary Classification: Normal vs. Diabetic Retinopathy.
Multi-class Classification: Five-level DR severity grading.

Feature Extraction Techniques:

Histogram of Oriented Gradients (HOG): Captures local shape and texture by computing the distribution of gradient orientations in small spatial cells. Each image is transformed into a high-dimensional vector of 26,244 features.
Topological Data Analysis (TDA): Uses cubical persistent homology to extract high-level descriptors. By applying sublevel filtrations across grayscale and RGB channels and vectorizing them using Betti functions ( $\beta_0$ for connected components and $\beta_1$ for loops/holes), the study extracts 800 topological features per image.

Machine Learning Framework:
The extracted features were independently evaluated using seven classical machine learning models: Logistic Regression, Random Forest, XGBoost, Support Vector Machine (SVM), Decision Tree, k-Nearest Neighbors (k-NN), and Extra Trees. All experiments used 10-fold cross-validation.

3. Key Contributions

First Comparative Study: Provides the first systematic head-to-head comparison between TDA and HOG in the context of medical image classification.
Feature Characterization: Demonstrates how HOG encodes local geometric cues while TDA encodes global structural connectivity.
Interpretability Analysis: Offers a mathematical explanation of how TDA features (Betti curves) can be visually interpreted to identify pathological changes (e.g., increased loop counts in DR images due to bright lesions).
Efficiency Insights: Shows that both methods are lightweight and do not require data augmentation or heavy pre-training.

4. Results

Binary Classification Performance:
- XGBoost was the top performer for both.
- HOG achieved slightly higher accuracy (94.29%) compared to TDA (94.18%).
- HOG showed superior performance in linear and kernel-based models (Logistic Regression and SVM).
Multi-class Classification Performance:
- The task is more complex, but results remained competitive.
- XGBoost achieved 74.41% with HOG and 74.69% with TDA.
- The performance gap between HOG and TDA narrowed in the multi-class setting, suggesting TDA is highly effective at capturing the structural nuances required to distinguish between disease severity levels.
Interpretability: HOG features were found to be difficult to map back to clinical anatomy, whereas TDA Betti curves provided clear, class-specific structural signatures.

5. Significance and Future Work

The study concludes that HOG and TDA are complementary. HOG is excellent for detecting local textural abnormalities, while TDA is superior for understanding the global organizational patterns of the retina.

Significance: This research provides a foundation for hybrid diagnostic systems. By integrating these handcrafted features into modern deep learning pipelines, researchers can create models that are not only highly accurate but also more interpretable and robust to noise.

Future Directions:

Integrating these features into end-to-end deep learning architectures.
Expanding the methodology to other medical modalities (MRI, CT, etc.).
Optimizing the computational scalability of TDA for real-time clinical deployment.