Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Whole Slide Image Classification of Salivary Gland Tumours

John Charlton, Ibrahim Alsanie, Syed Ali Khurram

TL;DR

This study tackles the challenge of classifying salivary gland tumours on whole-slide images (WSIs), a task made difficult by rarity and morphological diversity. It adopts a multiple instance learning (MIL) framework with CLAM for feature aggregation, comparing two patch-level feature extractors—ResNet-50 and CTransPath—and evaluating on a dataset of 646 WSIs for benign/malignant and adenoid cystic carcinoma classification. Results show that CTransPath paired with CLAM substantially outperforms ResNet-50, achieving AUROCs of $0.92$ for benign/malignant and $0.96$ for adenoid cystic carcinoma, with F1 scores of $0.88$ and $0.84$ respectively, aided by attention maps that localize diagnostically relevant regions. The findings demonstrate the viability of MIL on WSIs for SGTs and point to future work comparing against other architectures and expanding datasets to improve robustness and clinical utility.

Abstract

This work shows promising results using multiple instance learning on salivary gland tumours in classifying cancers on whole slide images. Utilising CTransPath as a patch-level feature extractor and CLAM as a feature aggregator, an F1 score of over 0.88 and AUROC of 0.92 are obtained for detecting cancer in whole slide images.

Whole Slide Image Classification of Salivary Gland Tumours

TL;DR

This study tackles the challenge of classifying salivary gland tumours on whole-slide images (WSIs), a task made difficult by rarity and morphological diversity. It adopts a multiple instance learning (MIL) framework with CLAM for feature aggregation, comparing two patch-level feature extractors—ResNet-50 and CTransPath—and evaluating on a dataset of 646 WSIs for benign/malignant and adenoid cystic carcinoma classification. Results show that CTransPath paired with CLAM substantially outperforms ResNet-50, achieving AUROCs of for benign/malignant and for adenoid cystic carcinoma, with F1 scores of and respectively, aided by attention maps that localize diagnostically relevant regions. The findings demonstrate the viability of MIL on WSIs for SGTs and point to future work comparing against other architectures and expanding datasets to improve robustness and clinical utility.

Abstract

This work shows promising results using multiple instance learning on salivary gland tumours in classifying cancers on whole slide images. Utilising CTransPath as a patch-level feature extractor and CLAM as a feature aggregator, an F1 score of over 0.88 and AUROC of 0.92 are obtained for detecting cancer in whole slide images.
Paper Structure (4 sections, 2 figures)

This paper contains 4 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Background and Methodology
  3. Results
  4. Disclosure of Interests.

Figures (2)

  • Figure 1: ROC curve of Benign/Malignant classification. The blue ROC curve is for ResNet-50 features. The orange curve is for CTransPath features.
  • Figure 2: Section of a whole slide image (WSI) showing the heatmap of attention. Areas highlighted in red are more important in deciding the categorisation of the whole image