Research in medical imagery field such as analysis of Red Blood Cells (RBCs) abnormalities can be used to assist laboratory’s in determining further medical actions. Convolutional Neural Networks (CNN) is a commonly used method for the classification of RBCs abnormalities in blood cells images. However, CNN requires large number of labeled training data. A classification of RBCs abnormalities in limited data is a challenge. In this research we explore a semi-supervised learning using Graph Convolutional Networks (GCN) to classify RBCs abnormalities with limited number of labeled sample images. The proposed method consists of 3 stages, i.e., extraction of Region of Interest (ROI) of RBCs from blood images using Faster R-CNN, abnormality labeling and abnormality classification using GCN. The experiment was conducted on a publicly accessible blood sample image dataset to compare classification performance of pretrained CNN models (Resnet-101 and VGG-16) and GCN models (Resnet-101 + GCN and VGG-16 + GCN). The experiment showed that the GCN model build on VGG-16 features (VGG-16 + GCN) produced the best accuracy of 95%.
Copyrights © 2022