Image Classification of Leukemic Cells Using Invariants of Triangle-Free Graphs as Synthetic Features

Development of efficient methods of the cellular image processing is an important avenue for practical application of modern artificial intelligence techniques. In particular, practical hematology requires automatic classification of images with or without leukemic (blast) cells in peripheral blood smears. This paper presents a new approach to the problem of classification of such cellular images based on graph theory, XGBoost algorithm and convolutional neural networks (CNN). Firstly, each image is transformed into a weighted graph using gradient of intensity. Secondly, a number of graph invariants are computed thus producing a set of synthetic features that is used to train machine learning model based on XGBoost. Combining XGBoost with CNN further increases the accuracy of leukemic cell classification. Sensitivity (TPR) and Specifity (TNR) of the XGBoost-based model were 95% and 97% accordingly; ResNet-50 model showed TPR of 95% and TNR of 98%. Combined use of the XGBoost-based and the ResNet-50 models demonstrated TPR of 99% and TNR of 99%.