Machine learning, unsupervised learning and stain normalization in digital nephropathology - Master thesis 2023

Aug 5, 2026

The thesis investigates how stain normalization impacts unsupervised learning clustering of glomeruli in digital nephropathology for chronic kidney disease. A three-stage unsupervised pipeline is implemented to compare clustering performance when applying Reinhard stain normalization, Macenko stain normalization, and structure preserving color normalization, using different reference images. Performance is quantified with the Adjusted Rand Index to measure agreement between clustering results and true labels. Results show increased cluster agreement for all normalization methods, with Reinhard achieving the strongest overall performance.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages62

Reading Time2 h 36 min

AccessFree to download

Summary

Introduction

Problem description

Aims and objectives

Background

Chronic kidney disease

Machine learning

Stain normalization

Methods

Overall approach

Data materials

Step 1: Stain normalization

Step 2: Image pre-processing and feature extraction

Step 3: Dimensionality reduction

Step 4: Clustering

Implementation

Available hardware

Image reading

Implementation of stain normalization techniques

Loading the pre-trained CNN and extracting features

Clustering

Results

No normalization

Reinhard

Macenko

SPCN

Discussion

Conclusions

Bibliography

Appendix

|

Faculty of Science and Technology Department of Electrical Engineering and Computer Science
MASTER THESIS | | | --- | --- | | Study line / specialization: Master in engineering / Signal processing and robotics with medical technology | Spring semester 2023
Open | | Author: Jon Tveit | | | Main supervisor Hrafn Weishaupt
Local supervisor Kjersti Engan
Co-supervisor: Sabine Leh | | | Title for thesis: Machine learning, unsupervised learning
and stain normalization in digital nephropathology | | | Study points: 30 | | | Subject words:
Digital Pathology
Stain normalization | Page numbers: 56
+ appendix: 2
Stavanger 15 . June 2023 | Contents Contents i Summary v 1 Introduction 1 1.1 Problem description ...................... 1 1.2 Aims and objectives ...................... 2 2 Background 4 2.1 Chronic kidney disease ..................... 4 2.1.1 The kidney biopsy and digital histopathology .... 5 2.1.2 Glomeruli and morphological changes ........ 5 2.1.3 Diagnosing CKD .................... 7 2.1.4 Stain variation ..................... 7 2.2 Machine learning ........................ 8 i 2.2.1 Supervised learning ................... 8 2.2.2 Unsupervised learning ................. 9 2.2.3 Dimensionality reduction and cluster agreement ... 11 2.3 Stain normalization ....................... 12 2.3.1 Reinhard stain normalization ............. 12 2.3.2 Macenko stain normalization ............. 12 2.3.3 Structure preserving stain normalization ....... 13 3 Methods 14 3.1 Overall approach ........................ 14 3.2 Data materials ......................... 15 3.3 Step 1: Stain normalization .................. 16 3.3.1 Reference images .................... 19 3.3.2 Reinhard stain normalization ............. 20 3.3.3 Macenko stain normalization ............. 21 3.3.4 Structure preserving colour normalization ...... 22 3.3.5 Masking tissue of tissue background ......... 23 3.3.6 Evaluating stain normalization results ........ 25 3.4 Step 2: Image pre-processing and feature extraction ..... 25 3.5 Step 3: Dimensionality reduction ............... 26 ii 3.6 Step 4: Clustering ....................... 26 3.6.1 Evaluating clustering results .............. 27 4 Implementation 29 4.1 Available hardware ....................... 29 4.2 Image reading .......................... 30 4.3 Implementation of stain normalization techniques ...... 30 4.3.1 Reinhard implementation ............... 30 4.3.2 Macenko implementation ............... 31 4.3.3 SPCN implementation ................. 31 4.3.4 Pipeline for normalizing images ............ 31 4.3.5 Structural similarity index metric ........... 32 4.4 Loading the pre-trained CNN and extracting features .... 32 4.5 Clustering ............................ 33 4.5.1 Dimensionality reduction implementation ...... 33 4.5.2 Clustering implementation ............... 33 4.5.3 Adjusted Rand Index .................. 33 5 Results 34 5.1 No normalization ........................ 36 5.2 Reinhard ............................ 36 iii 5.3 Macenko ............................. 38 5.4 SPCN .............................. 40 6 Discussion 43 7 Conclusions 45 Bibliography 50 Appendix I Appendix A .............................. I Appendix B .............................. IV iv Summary Chronic kidney disease is a serious health challenge and still, the field of study lacks awareness and funding. Improving the efficiency of diagnosing chronic disease is important. Machine learning can be used for various tasks in order to make CKD diagnosis more efficient. If the disease is discovered quickly it can be possible to reverse changes. In this project, we explore techniques that can improve clustering of glomeruli images. The current thesis evaluates the effects of applying stain normalization tonephropathological data in order to improve unsupervised learning clusterin

FAQs

What problem does the thesis address in digital nephropathology?

It addresses limitations in awareness and funding for chronic kidney disease and focuses on improving the efficiency of diagnosing related image patterns through better unsupervised clustering of glomeruli images.

Which stain normalization techniques are evaluated?

The thesis evaluates Reinhard stain normalization, Macenko stain normalization, and structure preserving color normalization, each tested with different reference images.

How is clustering performance measured?

Clustering performance is evaluated using the Adjusted Rand Index, comparing unsupervised clustering outcomes to true labels.

What is the main conclusion about the best method?

All three normalization techniques increase overall cluster agreement, and Reinhard stain normalization provides the best performance with the highest single-experiment and average Adjusted Rand Index scores.

Aug 5, 2026

Upload by Lucas Martin