Open AccessApplication of geostatistical functions and deep features to kidney biopsy images to differentiate focal segmental glomerulosclerosis from minimal change disease
Author(s) -
Justino Santos,
Romuere Silva,
Rodrigo Veras
Publication year - 2020
Language(s) - English
Resource type - Conference proceedings
DOI - 10.5753/sibgrapi.est.2020.12984
Subject(s) - minimal change disease , convolutional neural network , focal segmental glomerulosclerosis , random forest , medicine , artificial intelligence , pattern recognition (psychology) , classifier (uml) , kidney disease , pathological , computer science , pathology , radiology , kidney , glomerulonephritis
Chronic kidney diseases arise from acute or intermittent pathologies that have not been adequately treated, such as minimal change disease (MCD) and focal segmental glomerulosclerosis (FSGS). The accurate identification of these two diseases is of paramount importance, because their treatments and prognoses are different. Thus, we propose a method that is capable of differentiating MCD from FSGS based on images from pathological examinations. In the proposed method, we use four pre-trained convolutional neural networks and geostatistical functions to extract image features. Of the 8,720 extracted features, we selected 94 based on mutual information criteria, and in the classification step, we used a random forest classifier. The proposed method obtained an accuracy of 94.3% and Kappa index of 87.9%, a level that is regarded as “almost perfect”, confirming that our method is very promising.