Optimal Set of Features for Leukaemia Images with Extracted Areas of Interest

Abstract

Feature extraction was found effective in image classification across various studies. From a dataset containing leukemia images of four main categories 17 features were extracted including Haralick texture features, the size and number of white blood cells, and average colours. The process of feature extraction produced high accuracy when evaluated with some machine learning classifiers. Nonetheless, selective application of features was implemented on the modified dataset by extracting regions of interest (ROI). This approach resulted in 131,071 different combinations of features; some configurations achieved superior accuracy. This research aimed to identify the optimal combination of features within an expanded leukemia dataset through ROI extractions. Notably, cell size and count emerged as significant factors contributing to enhanced accuracy.