BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational Pathology

Amaya Gallagher-Syed, Henry Senior *, Omnia Alwazzan *, Elena Pontarini *, Costantino Pitzalis *, Myles J. Lewis *, Michael R. Barnes *, Luca Rossi, Gregory Slabaugh * · 2025

Read BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational Pathology by Amaya Gallagher-Syed, Henry Senior *, Omnia Alwazzan *, Elena Pontarini *, Costantino Pitzalis *, Myles J. Lewis *, Michael R. Barnes *, Luca Rossi, Gregory Slabaugh * (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-20880.txt.

Introduction

Amaya Gallagher-Syed and co-authors build BioX-CPath, a pathology AI that shows its reasoning in biological terms. Tissue slides stained in several colors hold rich clues, yet most models treat them as plain pictures. Doctors also distrust scores they cannot inspect.

The system turns each slide into a graph and reasons across stains jointly. A new pooling part, stain-aware attention, builds a patient summary that weights each stain by its worth. Scores top prior work on rheumatoid arthritis and Sjogren's disease slide sets. The readouts go further: attention per stain, spread gauges, and stain interplay scores show what swayed the call. Those views can be checked against known disease pathways. Code and guides are public for clinics and labs to test.

Catalog record

CREATORAmaya Gallagher-Syed, Henry Senior *, Omnia Alwazzan *, Elena Pontarini *, Costantino Pitzalis *, Myles J. Lewis *, Michael R. Barnes *, Luca Rossi, Gregory Slabaugh *
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-20880

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational Pathology Amaya Gallagher-Syed, Henry Senior *, Omnia Alwazzan *, Elena Pontarini *, Costantino Pitzalis *, Myles J. Lewis *, Michael R. Barnes *, Luca Rossi, Gregory Slabaugh * Abstract The development of biologically interpretable and explainable models remains a key challenge in computational pathology, particularly for multistain immunohistochemistry (IHC) analysis. We present BioX-CPath, an explainable graph neural network architecture for whole slide image (WSI) classification that leverages both spatial and semantic features across multiple stains. At its core, BioX-CPath introduces a novel Stain-Aware Attention Pooling (SAAP) module that generates biologically meaningful, stain-aware patient embeddings. Our approach achieves state-of-the-art performance on both Rheumatoid Arthritis and Sjogren’s Disease multistain datasets. Beyond performance metrics, BioX-CPath provides interpretable insights through stain attention scores, entropy measures, and stain interaction scores, that permit measuring model alignment with known pathological mechanisms. This biological grounding, combined with strong classification performance, makes BioX-CPath particularly suitable for clinical applications where interpretability is key. Source code and documentation can be found at: https://github.com/AmayaGS/BioX-CPath 0 0 footnotetext: Accepted for publication at CVPR 2025 . 1 Introduction Whole Slide Image (WSI) scanners capture high resolution, multi-magnification digital images of stained tissue biopsies presented on glass slides. The digitization of these biopsies has spurred the development of computational pathology methods. Analysis of these WSIs currently stands as one of the gold standard diagnostic and subtyping methods for many forms of cancers and autoimmune diseases, such as Rheumatoid Arthritis (RA) and Sjogren’s Disease. Different types of staining exist, which highlight different aspects of the tissue samples. Hematoxylin & Eosin (H&E) staining, a traditional and widely used technique, offers a broad view of tissue architecture and cellular morphology, with Hematoxylin staining cell nuclei a deep blue-purple, while Eosin stains cytoplasm and extracellular matrix in shades of pink. In contrast, Immunohistochemistry (IHC) is a more specialized technique that uses antibodies tagged with visual markers to identify specific proteins or cell types within tissue samples, allowing for precise localization and visualization of cell populations present in the tissue [ 42 ] . In cancer diagnostics, H&E staining remains the foundation for initial assessment and general diagnosis. However, IHC plays a crucial role in tumor classification, prognosis determination, and treatment selection by pinpointing specific cancer biomarkers, such Human Epidermal Growth Factor Receptor 2 (HER2), Estrogen Receptor (ER) and Progesterone Receptor (PR) [ 65 , 66 ] . For autoimmune diseases, while H&E staining identifies general patterns of inflammation and tissue damage, IHC becomes essential for a more nuanced understanding of the disease process. It highlights the types of immune cells present in inflammatory infiltrates, detects autoantibody deposits, and visualizes specific autoantigens targeted by the immune system [ 44 ] . In clinical pathology, a tissue sample will be taken and thinly sliced, and different stains applied to these slices, often with a reference H&E slide to verify tissue quality [ 42 ] . These multi-stain WSI stacks are rich in information about cellular types, tissue structures, and spatial patterns which relate to disease presentation and prognosis. Expert pathologists examining these stacks perform a semi-quantitative analysis, efficiently integrating information across both scale, stains, and images. Most state-of-the-art computational pathology methods so far have focused on H&E and the single-stain domain. Work that has tackled IHC has often done so in the context of cell quantification via cell segmentation [ 52 , 53 , 15 , 26 ] , as well as prediction and scoring of biomarkers [ 48 , 27 , 26 ] , or registration of multistain stacks [ 56 ] . Some methods have also explored the potential of H&E to IHC virtual staining techniques [ 49 , 59 , 38 ] or recently of using IHC as views for self-supervised representation learning [ 26 ] . Most of these approaches have concentrated on extracting information from IHC slides such that this could be predicted using H&E or on quantification of cell populations in IHC. This is because H&E is an older, more widely available and cost-effective technology. However, the use of IHC and more advanced techniques such as immunofluorescence is only set to grow in the coming years, associated with a decrease in technology cost and more advanced biomarker detection techniques [ 42 ] . There is therefore a clear need for methods which explicitly focus on integrating the complex cell landscapes across stains. To the best of our knowledge, few studies have concentrated on the issue of classification of unregistered, unannotated multistain datasets to date, with a single stain graph and mid/late fusion approaches developed in [ 13 ] and a multistain attention graph approach proposed in MUSTANG [ 17 ] . However, the value of computational pathology extends beyond classification tasks. Given the rich information contained in multi-stain data, these methods must be interpretable and generate actionable biological insights at both the disease and patient level. This interpretability serves two crucial purposes: it advances our understanding of underlying disease mechanisms, and it allows pathologists to verify that model predictions align with established biological knowledge, building trust in the system’s outputs. 1.1 Contributions 1. We introduce BioX-CPath, a biologically-driven graph-based model tailored to the complex cellular landscapes in multistain datasets. BioX-CPath works across multiple stains using semantic and spatial cues to capture complementary cellular and tissue information. 2. We propose a novel Stain-Aware Attention Pooling (SAAP) module that generates expressive, stain-aware patient embeddings. This module uniquely respects the biological and diagnostic diversity across stains, improving interpretability and diagnostic relevance. 3. We fully leverage the biological interpretability of BioX-CPath via derived metrics: stain attention and entropy scores, stain-stain interaction scores and Graph Neural Networks (GNNs) node heatmaps. These metrics provide detailed insights into stain relevance and inter-stain relationships, uncovering key biological patterns and interactions that contribute to disease pathology. 1.2 Related Work 1.2.1 Multiple Instance Learning WSIs are gigapixel, heterogeneous image files, which present challenges for computer vision methods given each image can reach over 100k\times 100k pixels, generally within a low patient sample setting. A weakly supervised Multiple Instance Learning (MIL) approach is most often employed to address this challenge. The image is divided into a regular grid of smaller patches (e.g., 224\times 224 pixels), each inheriting slide/patient labels. Patches are then embedded into a feature vector and classified at the slide/patient level using some form of non-trainable (e.g., max or mean) or trainable aggregation on the set of instances. Methods such as ABMIL [ 25 ] , DS-MIL [ 36 ] , and CLAM introduced trainable linear attention aggregation layers [ 39 ] . TransMIL [ 51 ] tackles the issue of long-range dependencies by approximating self-attention operations between patches via the Nyström method [ 57 ] . 1.2.2 GNNs in Histopathology Applications in histopathology can be divided into cell, patch, or tissue-level graphs, with both node and graph classification approaches being employed. Patch-graphs can be constructed using features extracted from a WSI or a set of WSIs, then connected via edges [ 2 , 3 , 63 , 64 ] . DeepGraphConv [ 37 ] , PatchGCN [ 9 ] , GTP [ 63 ] , CAMIL [ 16 ] and HEAT [ 8 ] adopt this approach by constructing a graph connecting either the k

Read the complete text