Classification of Histopathology Slides with Persistent Homology Convolutions

This paper introduces Persistent Homology Convolutions, a novel method that captures local topological features in histopathology slides, demonstrating that this approach outperforms standard CNNs in classification accuracy and hyperparameter robustness by effectively integrating geometric information into deep learning models.

The Gist

Imagine you are a detective trying to solve a mystery: Is this tissue sample healthy, or is it cancerous?

In the world of medicine, pathologists (the detectives) look at tiny slides of tissue under a microscope. They don't just look at the colors; they look at the shape and arrangement of the cells. Are the cells crowded together? Are there holes in the tissue? Are the nuclei (the cell's "brain") split into multiple pieces? These geometric clues are vital for diagnosing diseases like Osteosarcoma (bone cancer).

For a long time, computers have tried to help with this using AI (Artificial Intelligence), specifically a type called Convolutional Neural Networks (CNNs). Think of a CNN as a very smart robot that scans an image, looking for patterns.

The Problem: The Robot is "Topologically Blind"

Here is the catch: Standard AI robots are great at seeing colors and edges, but they are terrible at understanding shape and structure.

Imagine you have a photo of a crowd of people.

• Standard AI: Counts the people. It sees "lots of heads."
• The Missing Piece: It doesn't understand how they are standing. Are they holding hands in a circle? Are they standing in a chaotic pile? Are there empty spaces between them?

In medical terms, standard AI often misses the "topology" (the study of shapes and holes). If a tumor causes cells to cluster in a weird, specific way, a standard AI might miss it because it's too focused on individual pixels rather than the big picture of the shape.

The Old Solution: The "Global" Summary

Some researchers tried to fix this by giving the AI a "global summary" of the shape.

• The Analogy: Imagine trying to describe a city to someone who has never seen it. You give them a single statistic: "This city has 500 parks and 1,000 buildings."
• The Flaw: This tells you how many things there are, but not where they are. A city with parks scattered everywhere feels very different from a city where all the parks are in one giant block. The "global summary" loses the local details that matter.

The New Solution: Persistent Homology Convolutions (PHC)

The authors of this paper invented a new tool called Persistent Homology Convolutions (PHC).

Think of PHC as a smart, sliding magnifying glass that doesn't just look at pixels, but looks at shapes and holes in real-time.

1. The Sliding Window: Instead of looking at the whole image at once, the AI slides a small window (like a 32x32 pixel square) across the entire slide, just like a person scanning a document with their eyes.
2. The "Shape Detective": Inside that small window, the AI doesn't just count pixels. It asks:
   • "How many separate groups of cells are here?"
   • "Are there holes (voids) between the cells?"
   • "Do these holes get bigger or smaller as we zoom out?"
3. The "Fingerprint": For every little window, the AI creates a tiny "fingerprint" (a vector) that describes the shape of that specific area.
4. The Assembly: It then stitches all these tiny fingerprints together to build a complete map of the tissue's geometry.

Why is this better?

• It keeps the "Local" context: It knows that a hole in the top-left corner is different from a hole in the bottom-right. It preserves the arrangement of the cells.
• It's translation invariant: If you shift the image slightly, the AI still recognizes the same shapes.
• It's efficient: Because it summarizes the shape into a simple "fingerprint" before feeding it to the main AI, the computer doesn't have to crunch as much raw data. It's like summarizing a 500-page book into a 10-page outline before reading it.

The Results: The Detective Wins

The researchers tested this new method on a dataset of bone cancer slides. They compared three approaches:

1. Standard AI: Just looking at the raw image.
2. Old Topology AI: Looking at the whole image's shape summary (the "global" method).
3. New PHC AI: Using the sliding window shape detective.

The Winner: The New PHC AI won hands down.

• It achieved 93.9% accuracy (compared to about 91% for standard methods).
• It was more consistent and didn't get confused by small changes in how the data was set up.
• It was faster to compute than the old "global" method.

The Big Takeaway

This paper is like upgrading a detective's toolkit. Instead of just giving the detective a list of clues (pixels) or a single summary of the crime scene (global shape), they gave the detective a magnifying glass that highlights the structure of the evidence as they walk through the scene.

By teaching the computer to understand the geometry and arrangement of cells locally, rather than just their colors, we can build AI that diagnoses cancer more accurately, potentially saving more lives.

Technical Summary

Here is a detailed technical summary of the paper "Classification of Histopathology Slides with Persistent Homology Convolutions" by Pothagoni and Schweinhart.

1. Problem Statement

Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) are standard tools for medical image classification, often achieving accuracy comparable to pathologists. However, these architectures have a critical limitation: they tend to lose topological information during processing.

• The Issue: CNN pooling operators and ViT patch subdivisions alter the geometric structure of the image. In histopathology, the geometric arrangement of cells (e.g., cell size, multinucleation, tissue disorganization) is a primary indicator of disease (such as Osteosarcoma).
• The Gap: While previous research has attempted to integrate Persistent Homology (PH)—a tool from topological data analysis that quantifies shape features like connected components and holes—into machine learning, existing methods compute PH globally (across the entire image). Global summaries fail to capture the locality of topological features, which is crucial for distinguishing between different tissue states (e.g., viable tumor vs. necrotic tumor).

2. Methodology

The authors propose a novel operator called Persistent Homology Convolutions (PHC) to reintroduce local topological information into CNNs.

A. Persistent Homology Convolutions (PHC)

Instead of computing PH on the whole image, PHC applies a sliding window approach similar to standard CNN convolutions but operates on topological data.

• Mechanism: The image is divided into overlapping sub-windows (patches) of size $M \times M$ with a stride $c$.
• Process:
  1. For each window, a filtration is applied to the pixel data (either an Alpha Complex based on thresholded pixels or an Extended Lower Star Filtration based on grayscale values).
  2. Persistent Homology is computed for that specific window, generating a Persistence Diagram (PD) containing intervals $(b, d)$ representing the birth and death of topological features.
  3. The PD is vectorized into a fixed-dimensional Persistence Image (using the method by Adams et al., 2017).
  4. These local persistence images are arranged into a 3D array and convolved with a learnable kernel matrix $K$.
• Advantage: This preserves translation equivariance and captures the relative placement of topological features, which global methods miss.

B. Filtration Types

The study utilizes two primary filtrations to generate the local PH data:

1. Alpha Complex: Derived from thresholded images (binary), focusing on the geometry of cell boundaries and voids.
2. Extended Lower Star / Adjacency Complex: Derived from grayscale intensity values, capturing features based on pixel intensity gradients (e.g., dark nuclei vs. lighter cytoplasm).

C. Experimental Setup

• Dataset: Osteosarcoma histopathology slides (1,144 images) from the Cancer Imaging Archive, classified into three classes: Non-Tumor, Non-Viable (Necrotic) Tumor, and Viable Tumor.
• Preprocessing: Images were resized to $512 \times 512$, converted to grayscale, and conditioned via thresholding/truncation and erosion to emphasize tissue architecture.
• Model Architecture: A lightweight CNN with two convolution/pooling layers followed by dense layers. The authors tested models trained on:
  1. Raw Grayscale Images.
  2. Global Persistent Homology (PH).
  3. Local PH (PHC).
  4. Hybrid models (Images + PHC or Images + Global PH).
• Evaluation: Over 10,000 experiments were conducted using Bayesian hyperparameter sweeps to optimize for accuracy, precision, sensitivity, and specificity.

3. Key Contributions

1. Mathematical Definition of PHC: The paper formally defines the Persistent Homology Convolution operator, bridging the gap between topological data analysis and deep learning convolution operations.
2. Local vs. Global Topology: It demonstrates that local topological summaries are superior to global summaries for histopathology, as they retain critical spatial relationships between cellular structures.
3. Comprehensive Empirical Study: The authors conducted a massive comparative study (10,000+ runs) across multiple filtration types, dimensions (0D and 1D), and data combinations.
4. Open Source: The authors released a public repository containing the PHC implementation and experimental setup.

4. Results

The models trained with PHC consistently outperformed conventional CNNs and global PH approaches across all metrics.

• Accuracy: The best-performing model (Hybrid: Images + Alpha Complex PHC) achieved 93.9% accuracy (Table 1). This surpasses the previous state-of-the-art on this dataset (91.2% with standard CNNs) and significantly outperforms global PH methods (74.2%).
• Robustness: Models trained with PHC showed lower variance and were less sensitive to hyperparameter choices compared to standard image-based models.
• Intrinsic Dimensionality: Analysis using Maximum Likelihood Estimators (Table 3) revealed that PHC data has a substantially lower intrinsic dimension than raw images (e.g., 4.41 vs. 12.50 for Alpha Complex PHC). This suggests PHC effectively reduces data complexity while retaining biologically relevant geometric information.
• Efficiency: PHC computation is significantly faster than global PH. For a single image, PHC took 2.6 seconds (Alpha Complex) compared to 6.8 seconds for global PH, and 11.4 seconds vs. 3496 seconds for Extended Adjacency Complex (Table 2).
• Feature Importance: The study found that 1-dimensional persistence features (loops/holes) were generally more predictive than 0-dimensional features (connected components), indicating that the geometry of empty space and cell boundaries is critical for classification.

5. Significance

• Improved Diagnostics: The method provides a more accurate tool for distinguishing between tumor viability states, which is critical for treatment planning in Osteosarcoma.
• Topological Deep Learning: This work validates the hypothesis that local topology is a distinct and necessary feature for medical image analysis that standard CNNs fail to extract automatically.
• Efficiency: By reducing the intrinsic dimensionality of the data, PHC allows for smaller, faster, and more robust models that require fewer parameters to achieve high accuracy.
• Generalizability: The approach is not limited to Osteosarcoma; the framework of PHC can be applied to any domain where local geometric structure is a key differentiator (e.g., materials science, other tissue types).

In conclusion, the paper successfully demonstrates that integrating local persistent homology via a convolution-like operator significantly enhances the performance and robustness of deep learning models in histopathology, offering a new paradigm for geometric deep learning in medicine.