A Comparison of Convolutional Neural Networks, Image Feature Extraction, and Clinical Features Applied to Leukemia Blood Smear Image Classification
Marinela Branescu, Steve Counsell, Allan Tucker, Stephen Swift · Artificial intelligence · 2025
The study presents the performance of Convolutional Neural Network (CNN) and Standard Feature extraction for image classification in detecting the type of leukaemia from blood smear test images. This research uses 312 pre-classified leukaemia images belonging to the four main types of the disease: Chronic Myeloid Leukaemia (CML), Chronic Lymphatic Leukaemia (CLL), Acute Myeloid Leukaemia (AML), and Acute Lymphatic Leukaemia (ALL). Using CNN trained by standard packages and traditional feature extraction with machine learning classifiers to classify the leukaemia images, the modest accuracy for the first experiment and longer implementation time for the second experiment leads to using image manipulation techniques. The Otsu thresholding method with respect to the adapted Otsu 2-3 method is efficient in accentuating the cell’s shape preparing the images for extracting regions of interest (ROI). ROI dataset is further filtered of residues expressed in unclear parts of white blood cells, obtaining an ROI-filtered dataset. The obtained separated datasets following the application of image-manipulated techniques are evaluated separately with both methods and the changes in accuracy and efficiency of the system are contrasted. ROI dataset improves the CNN accuracy by increasing the number of images and reducing unnecessary pixels. Still, feature extraction performance is altered and selective feature extraction is applied, some selections of features producing higher accuracy. A reduced dataset of images will impact CNN’s performance, and the ROI dataset will perform better. Feature extraction is efficient, but a lengthier process and selective application of features are necessary to validate the importance of some features over others.