📄 other

Hidden networks, deconstructions, convolutions and colourful revolutions. Explainable multi- class classification for histopathologic lymphoid tissue prototyping

This study presents an explainable, unsupervised multi-class classification framework using Minkowski and Cauchy kernel density methods to achieve 95.7% accuracy in prototyping CD20-stained lymphoid tissue samples, while highlighting unusual diaminobenzidine signaling patterns and identifying the need for expanded immunohistochemical and genetic data to distinguish B and T cell lineages.

Original authors: Arshaad Sibda, Turgay Celik, Reubina Wadee

Published 2026-07-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Arshaad Sibda, Turgay Celik, Reubina Wadee

Original paper licensed under CC BY 4.0 (https://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 solve a mystery inside a tiny, crowded city made of cells. In the world of medicine, this city is a slice of tissue from a patient's body, viewed under a microscope. Usually, doctors (pathologists) look at these cities with their eyes, hunting for clues in the shapes and colors of the buildings (cells) to figure out what kind of disease is invading. But sometimes, the clues are tricky, the lighting is weird, or the city is just too big to look at every single brick. This is where "deep learning" comes in. Think of deep learning as a super-smart, tireless robot apprentice that can scan millions of these tiny cities in seconds. However, there's a catch: most of these robots are "black boxes." They give you an answer like "It's a B-cell problem!" but they can't explain why. They don't show you their work. This paper tackles that mystery by building a robot that doesn't just guess, but actually shows its detective board, explaining exactly which clues led to the conclusion. It's about making the invisible logic of AI visible and understandable for the human doctors who need to trust it.

The researchers behind this study, working at the University of the Witwatersrand in South Africa, decided to build a new kind of "detective robot" specifically for looking at lymphoid tissue (a type of immune tissue) that has been stained with a special brown dye called CD20. This dye acts like a highlighter, marking specific B-cells (a type of white blood cell) so they stand out against the blue background of other cells. Their goal was to create a system that could not only sort these tissues into different categories with high accuracy but also "deconstruct" its own thinking process so a human could see exactly how it reached a decision.

To do this, they didn't just feed the computer a picture and ask for an answer. Instead, they built a three-lane highway for the data, where the image of the tissue travels through three different processing streams simultaneously, like a car taking three different routes to the same destination to compare notes.

Stream A is like a shape-and-color scanner. It breaks the image down into tiny grids to count the edges and shapes (using something called HOG) and also takes a snapshot of the exact mix of red, green, and blue colors (RGB) in the tissue.
Stream B is the texture detective. It uses a toolkit of advanced math tricks to look at the "grain" of the tissue. It checks for patterns in how the colors repeat (LBP), measures how the light and dark spots interact (Haralick), and even uses a wave-like math tool (Wavelets) to find hidden details in the texture. It also specifically looks for the brown dye (DAB) to measure exactly how much of the target cells are present.
Stream C is a small, fast neural network (a mini-brain) that looks at the image as a whole to catch the "big picture" structures that the other two streams might miss.

Once these three streams gathered their clues, the system didn't just throw them into a pile. It fused them together into a massive, 6,657-dimensional "super-feature" vector. To make this manageable, the researchers projected this giant data cloud onto a "unit hypersphere." Imagine a giant, invisible ball where every possible tissue pattern is a dot on the surface. The system then uses two different scoring methods—Minkowski (MK) and Cauchy Kernel Density (CKD)—to see how close a new tissue sample is to the "prototypes" (the average examples) it has already learned. It's like checking if a new suspect looks more like the "B-cell gang" or the "T-cell gang" based on how close they stand to the group leaders on this giant ball.

The results were impressive. The system achieved a classification accuracy of 95.7% on a validation set of 303 tiles (small pieces of the tissue image). It correctly identified 91 tiles of Sample 1, 88 of Sample 2, 14 of Sample 3, and 93 of Sample 4. The researchers found that the two different scoring methods (MK and CKD) gave identical results, suggesting the system's success came from the quality of the clues it gathered, not just the math used to count them.

One of the most fascinating discoveries happened with "Sample 1." The system spotted something unusual: the brown dye (CD20) was present but weak and patchy. The researchers noted this was likely due to the cells being in a specific state where they had stopped making the CD20 protein, possibly because of previous treatments or a suppressed immune system. The system didn't just say "wrong"; it highlighted this "unusual diaminobenzidine signalling" as a key feature, showing that it could spot subtle biological quirks that a simple "yes/no" classifier might miss.

However, the paper is careful not to claim this is a magic bullet that solves everything. The authors explicitly state that their model relies on a single stain (CD20) and that it struggled slightly with Sample 3, which had fewer data points to learn from (only 52 original tiles, augmented to 208). They also admit that the model couldn't perfectly distinguish between B-cells and T-cells in all cases because it lacked a second stain (like CD3) to confirm the T-cell presence. In fact, they note that three B-cell dominated tiles were mistakenly classified as Sample 4 (which was mostly T-cells), suggesting that without more genetic or multi-stain data, the robot can still get confused at the boundaries.

The study concludes that while the system is highly effective at its specific task—sorting these four types of lymphoid tissue with 95.7% accuracy and providing a clear visual map of why it made those choices—it is not yet a standalone diagnostic tool for all diseases. The authors suggest that future work needs to include a full panel of stains and genetic data to help the robot distinguish between different types of cell lineages more clearly. They have built a transparent, explainable framework that works well on the data it was given, but they caution that generalizing this to the wider, messier world of human disease will require more training and more complex clues.

Technical Summary: Explainable Multi-Class Classification for Histopathologic Lymphoid Tissue Prototyping

Problem Statement
The classification of chromogen-rich, feature-rich histopathologic lymphoid tissue typically involves the dimensional reduction of gigapixel whole slide images (WSIs) and the deconstruction of tile-level data into sub-space feature components. These processes often rely on deep learning neural networks (DNNs) that function as "black boxes," offering high performance but lacking mechanistic interpretability for pathologists. Furthermore, assessment ambiguity frequently arises under brightfield microscopy, and existing frameworks often lack sufficient visualisation to bridge the gap between abstract feature extraction and human understanding. This study addresses the need for an explainable, unsupervised multi-class classification framework capable of accurately classifying CD20 immunohistochemically (IHC) stained lymphoid samples while visualising the operations of the model for interpretability.

Methodology
The study employed a 3-stream feature-extractor backbone engineered on a CPU-based local server, processing CD20 IHC stained lymphoid tissue samples dominated by B lymphocytes. The methodology proceeded through the following stages:

Data Collection and Pre-processing:
- Samples were derived from patients at two academic hospitals in South Africa, covering periods from 1990–2007 and 2017–2020.
- Four distinct glass slides (Samples S1–S4) were selected, anonymised, and scanned at 20x magnification using a Leica Biosystems Aperio AT2 scanner.
- Pre-processing involved removing noise (dust, folds, air bubbles) and extracting non-overlapping 512x512 pixel tiles. Sample S3, initially containing only 52 usable tiles, was synthetically augmented to 208 tiles to address statistical imbalance.
- The final dataset comprised 652 training tiles and 303 validation tiles.
Three-Stream Feature Extraction:
The model extracted features through three parallel streams, each normalised independently before fusion:
- Stream A (SA): Combined Histogram of Oriented Gradients (HOG) with an RGB colour histogram. This stream captured edge-shape information and spectral characteristics, resulting in a 6,468-dimensional vector.
- Stream B (SB): A composite of Haralick Cell Contrast (HGLCM), Local Binary Patterns (LBP), Discrete Wavelet Transform (DWT), Watershed Cell Morphometry (WCM), and CD20 DAB characterisation. This stream extracted textural, microstructural, and morphometric features, resulting in a 61-dimensional vector.
- Stream C (SC): A ShallowNet Convolutional Neural Network (SN-CNN), a four-layer encoder processing 96x96 RGB tiles to produce a 128-dimensional feature vector. This provided coarse spatial structural regularisation.
Feature Fusion and Hyperspherical Embedding:
- Features from the three streams were concatenated and projected onto a unit hypersphere to form a 6,657-dimensional unit-hypersphere representation ( $\phi(T)$ ).
- Scaling hyperparameters ( $\alpha_{sb}=0.18$ and $\alpha_{sn-cnn}=0.05$ ) were applied to control the angular influence of Streams B and C relative to the Stream A backbone, ensuring the fused vector remained scale-invariant.
Classification and Learning Procedure:
- The study utilised an unsupervised prototype learning approach adapted from Angelov's work on autonomous mobility-based image prototyping.
- Two scoring-based classifiers were implemented: Minkowski (MK) and Cauchy Kernel Density (CKD).
- The algorithm incrementally learned prototypes on the unit hypersphere. New tiles were either absorbed into existing prototypes (if within a calibrated radius $r_0 \approx 0.134$ ) or created as new prototypes.
- Classification decisions were based on the nearest prototype distance, converted to a score via an exponential function.

Key Results

Performance: Both the MK and CKD classifiers achieved identical high-performance metrics on the validation set: 95.7% accuracy, a Cohen's kappa score of 0.938, and a macro-F1 score of 95.1%.
Prototype Generation: From 808 augmented tiles, the model derived 335 unique prototypes. The compression ratios varied by sample, with S3 showing the highest compression (3.15x), indicating high internal homogeneity.
Interpretability and Visualisation:
- PCA Analysis: Principal Component Analysis (PCA) of the 6,657-dimensional space revealed that S1 was the most linearly separable, while S2 was the most heterogeneous and difficult to classify, straddling the decision boundaries of other samples.
- DAB Signalling: The model successfully identified unusual diaminobenzidine (DAB) signalling in Sample 1, attributing it to a high likelihood of immunosuppression and a dominant plasmablast-like population where CD20 is typically downregulated.
- Boundary Analysis: Confusion matrix decomposition and Voronoi tessellation visualisations confirmed that misclassifications were concentrated at decision boundaries (specifically between S2 and S4) rather than distributed across the feature space, validating that errors were due to feature representation limits rather than scoring mechanism failures.

Significance and Claims
The paper claims to address the paucity of explainable classification frameworks that provide abundant visualisations for pathologists. By integrating interpretability constraints, the study bridges the trade-off between "glass-box" (mechanistic) and "black-box" approaches. The authors assert that their framework successfully:

Realises targeted, high-quality feature extraction for multi-class IHC stain modalities.
Provides domain-specific morphometric feature depth at 20x magnification.
Maintains explainability and interpretability without impeding feature fusion.
Operates efficiently using CPU-based computing resources.

The study concludes that while the model achieved high accuracy on a single IHC stain dataset, its generalisation to disease class levels beyond these constraints is limited. The authors explicitly note that the exclusion of CD79a (due to unknown rituximab history) and the lack of T-cell lineage markers (CD3) or genetic translocation data are major limitations. Future investigations are directed toward incorporating complete immunohistochemistry panels and genetic labels to discriminate between B-cell and T-cell lineage lesions.

Technical Summary: Explainable Multi-Class Classification for Histopathologic Lymphoid Tissue Prototyping

More like this