White Blood Cell Classification of Porcine Blood Smear Images

Abstract

Differentiating white blood cells has been a fundamental part of medical diagnosis as it allows the assessment of the state of health of various organ systems in an animal. However, the examination of blood smears is time-consuming and is dependent on the level of the health professional’s expertise. With this, automated computer-based systems have been developed to reduce the time taken for examination and to reduce human error. In this work, an image processing technique was explored to investigate the classification of white blood cells. Through this technique, color and shape features were gathered from segmented nuclei and cytoplasms. Various deep learning algorithms where transfer learning methods were also employed for comparison. Experimental results showed that handcrafted features via image processing are better than features extracted from pre-trained CNNs, achieving an accuracy of 91% when using a non-linear SVM classifier. However overall, deep neural networks were superior in WBC classification as the fine-tuned DenseNet-169 model was found to have the highest accuracy of 93% against all used methods.