Machine learning, unsupervised learning and stain normalization in digital nephropathology

Abstract

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 to nephropathological data in order to improve unsupervised learning cluster- ing. A unsupervised learning pipeline was implemented in order to evaluate the effects of using stain normalization techniques with different reference images. The stain normalization techniques that were implemented are: Reinhard stain normalization, Macenko stain normalization and Structure preserving color normalization. The evaluation of these methods was done by measuring clustering results from the unsupervised learning pipeline, using the Adjusted Rand Index metric. The results indicate that using these techniques will increase the cluster agreement between results and true labels for the data. Six reference images were used for each stain nor- malization technique. The average Adjusted Rand Index score for all ref- erence images was increased using all three stain normalization techniques. The best performing method overall was the Reinhard stain normalization technique. This method gave both the highest single experiment and aver- age score. The other normalization methods both have one score close to zero (unsuccessful clustering), and structure preserving color normalization would outperform the Reinhard method if this single clustering was more successful.

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 to nephropathological data in order to improve unsupervised learning cluster- ing. A unsupervised learning pipeline was implemented in order to evaluate the effects of using stain normalization techniques with different reference images. The stain normalization techniques that were implemented are: Reinhard stain normalization, Macenko stain normalization and Structure preserving color normalization. The evaluation of these methods was done by measuring clustering results from the unsupervised learning pipeline, using the Adjusted Rand Index metric. The results indicate that using these techniques will increase the cluster agreement between results and true labels for the data. Six reference images were used for each stain nor- malization technique. The average Adjusted Rand Index score for all ref- erence images was increased using all three stain normalization techniques. The best performing method overall was the Reinhard stain normalization technique. This method gave both the highest single experiment and aver- age score. The other normalization methods both have one score close to zero (unsuccessful clustering), and structure preserving color normalization would outperform the Reinhard method if this single clustering was more successful.