Classificação de Imagens de Biópsias Renais com Glomeruloesclerose Segmentar e Focal ou com Lesões Mı́nimas Utilizando Transfer Learning em CNN

Chronic renal diseases arise from acute or intermittent, not adequa- tely treated pathologies such as minimal change disease (MCD) and focal seg- mental glomerulosclerosis (FSGS). Correct identification of these two diseases is of paramount importance because their treatments and prognoses are diffe- rent. Thus, we propose a method capable of differentiating MCD and FSGS through images of pathological exams. In the proposed method, we extracted 10240 features from three pre-trained convolutional neural networks, we se- lected 62 from them through the mutual information algorithm, and we used the Random Forest for the classification. The method obtained an accuracy of 93.33% and Kappa of 85.47%, which is considered “Almost Perfect”.