TVT-PAPD: Pathology-Aware Prototype Distillation for Self-Supervised Whole Slide Image Classification

Jul 16, 2026

Self-supervised learning (SSL) enables transferable representations from large unlabeled whole slide images (WSIs), yet many methods primarily learn generic visual features and miss pathology-specific morphological patterns essential for disease characterization. TVT-PAPD introduces a Tiny Vision Transformer (TVT) coupled with Pathology-Aware Prototype Distillation (PAPD), using a learnable pathology prototype bank to preserve representative tissue morphology and align consistent discriminative representations. Experiments on TCGA LGG/GBM and IPD-Brain report weighted F1-scores of 93.02% and 90.23%, with strong cross-cohort generalization and improved interpretability.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

Introduction

Digital pathology and WSI challenges

Self-supervised learning and pathology foundation models

Methodology

Algorithm: Tissue patch selection for TVT-PAPD

arXiv :2607 . 10406v1 [ cs .CV] 11 Jul 2026 TVT-PAPD: PATHOLOGY-AWARE PROTOTYPE DISTILLATION FOR SELF-SUPERVISED WHOLE SLIDE IMAGE CLASSIFICATION Ramesh Naidu Laveti1 , Jaya Sreevalsan-Nair1 T K Srikanth 1E-Health Research Center, International Institute of Information Technology Bangalore, Karnataka 560100, India. [https://ehrc.iiitb.ac.in/](https://ehrc.iiitb.ac.in/) July 10, 2026 ABSTRACT Self-supervised learning (SSL) has emerged as an effective paradigm for learning transferable representations from large-scale unlabeled whole slide images (WSIs) . However, existing SSL methods primarily learn generic visual features and often fail to explicitly capture pathology-specific morphological patterns that are critical for disease characterization. To address this limitation, we propose Tiny Vision Transformer with Pathology-Aware Prototype Distillation (TVT-PAPD) . This self-supervised pathology representation learning framework integrates a Tiny Vision Transformer (TVT) with a novel Pathology-Aware Prototype Distillation (PAPD) module. PAPD employs a learnable pathology prototype bank to discover and preserve representative tissue morphology patterns, encouraging semantically similar pathological regions to learn consistent and discriminative representations. The proposed framework enhances pathology-aware feature learning while maintaining computational efficiency with 90M parameters. Experiments on the Cancer Genome Atlas (TCGA) low-grade glioma (LGG)/glioblastoma (GBM) dataset and the Indian Pathology Brain (IPD-Brain) dataset demonstrate that TVT-PAPD achieves weighted F1-scores of 93.02% and 90.23%, respectively, for LGG-GBM classification, while exhibiting strong cross-cohort generalization across independent glioma datasets. Furthermore, the learned prototypes capture meaningful patterns in tissue morphology, improving the interpretability of the representation. These results highlight the effectiveness of pathology-aware prototype distillation for efficient, robust, and interpretable WSI analysis in computational pathology. Keywords Computational Pathology, Whole Slide Images, Self-Supervised Learning, Vision Transformers, Pathology-Aware Prototype Distillation, Foundation Models, Histopathology Image Analysis, Glioma Classification. 1 Introduction Digital pathology has transformed computational pathology by enabling the digitization of histopathological slides into high-resolution WSIs [1] . These gigapixel-scale images contain rich morphological information for automated diagnosis, prognosis, and disease screening [2, 3] . However, their massive size and tissue heterogeneity make efficient WSI analysis challenging [4] . Furthermore, staining variability and limited expert annotations further complicate robust model development [5, 6] . Consequently, WSIs are typically processed as collections of image patches [4] . Transformerbased architectures have substantially improved WSI analysis by modeling long-range contextual dependencies beyond the limited receptive fields of convolutional neural networks (CNNs) . Recent approaches, including Vision Transformer (ViT) [7], Clustering-constrained Attention Multiple Instance Learning (CLAM) [8], Transformer-based Multiple Instance Learning (TransMIL) [9], and Hierarchical Image Pyramid Transformer (HIPT) [10], have achieved strong performance in pathology image analysis, while pathology foundation models such as GigaPath [11] and Virchow [12] have demonstrated the benefits of large-scale pretraining. Nevertheless, these models often require considerable computational resources. ∗ [jnair@iiitb.ac.in](jnair@iiitb.ac.in) . JULY 10, 2026 Algorithm 1 Tissue Patch Selection for TVT-PAPD Require: Whole slide image I, patch size P, background threshold τ Ensure: Tissue patch set P 1: Initialize retained patch set P ← 0/ 2: for each grid location (x, y) in I with stride P do 3: t ← CropPatch (I, x, y, P) 4: ρ ← BackgroundRatio (t ) 5: if ρ

FAQs

What problem does TVT-PAPD address in existing self-supervised WSI learning?

Existing SSL methods often learn generic visual features and fail to explicitly capture pathology-specific morphological patterns needed for diagnostically meaningful tissue characterization. TVT-PAPD targets this by incorporating pathology-aware distillation.

How does Pathology-Aware Prototype Distillation (PAPD) work?

PAPD uses a learnable pathology prototype bank to discover and preserve representative tissue morphology patterns. During self-supervised training, teacher–student prototype distributions are aligned to encourage consistent and discriminative representations.

What performance results are reported for glioma classification?

On TCGA LGG/GBM, TVT-PAPD achieves a weighted F1-score of 93.02%, and on the IPD-Brain dataset it achieves 90.23%. The results also indicate strong cross-cohort generalization and improved interpretability from learned prototypes.

Jul 16, 2026

Upload by Theodore