A Machine Learning Algorithm Segments Renal Tissue into Healthy and Pathological Structures
Researchers from the Netherlands create a method using convolutional neural networks to automate kidney sample segmentation.
Typically, kidney disease is diagnosed by manually inspecting histopathological slides of renal tissue with a microscope. Recent studies have explored the idea of automated assessment of histopathology using the combination of high-resolution, whole-slide image scanners and machine learning approaches.
To potentially reduce the workload of pathologists, a team of researchers from Radboud University Medical Center in the Netherlands has created an algorithm that can partition an image of a kidney specimen into segments corresponding to different classes of tissue, a process known as segmentation. It employs convolutional neural networks, a type of machine learning that has successfully been applied to image recognition and classification. The results were presented in February during the 2018 SPIE Medical Imaging Conference in Houston, Texas.
The tissue samples were obtained from nine patients who underwent kidney removal. A renal histopathology expert manually annotated the slides in order to validate the accuracy of the machine learning models. The segmentation algorithms then divided the renal tissue into nine different classes: five present in healthy tissue, including glomeruli, arterioles and capillaries, and four seen in pathological tissue, including sclerotic glomeruli, inflammatory infiltrate and fibrotic tissue.
The authors reported on the accuracy of the machine learning models by comparing the manual annotations with the algorithm's results. Each pixel was labeled as either being correctly or incorrectly identified by the model. The best-performing model correctly identified 90 percent of five out of nine classes. All models seemed to have difficulties segmenting capillaries and arterioles due to their small size and low number of annotations, respectively. Although the work remains in early stages, the authors hope to create an automated method for renal tissue segmentation that can support both researchers and clinicians.