Lightweight Classification of Normal Versus Leukemic Cells Using Feature Extraction

Current deep learning approaches, based on convolutional neural networks, produce models with excellent performance in lymphocyte classification tasks. However, they require long training times and specific hardware to work. To overcome these drawbacks, we present a lightweight image classification system capable of discriminating between cancerous and noncancerous lymphocytes of leukemia patients using image processing and deep learning techniques that require much less training time and can be run on a standard CPU. Feature vectors are built with statistical, textural, morphological, contour, and frequency features extracted from each lymphocyte image, and used to train a neural network that classifies the lymphocytes into malignant or healthy. The proposed method has lower computational cost than other deep learning methods in both learning time and neural network size. Our results contribute towards leukemia classification systems showing that high accuracy can be achieved by a neural network classifier trained with a rich set of features.