White Blood Cells Classification Using Built-in Customizable Trained Convolutional Neural Network

This paper presents a new method of applying Convolutional Neural Network (CNN) on microscopic blood smear images to classify leukocytes or commonly known as white blood cells (WBCs). The proposed classification method leverages an existing novel segmentation algorithm to extract the template or sample WBC images from a given blood smear image. These template images are used to train the built-in customizable CNN, which is later simulated to classify each sample image as one of the five existing types of leukocytes. A total of ten blood smear images, each with more than one type of WBC, were used in our experiments. Furthermore, we increased the recognition rate of our proposed approach incrementally by training five unique CNNs. We were able to identify an optimum trained CNN out of the five trained CNNs and achieved an average recognition rate of 84%. Using our built-in custom trained CNN, we have been able to not only achieve a competitive recognition rate with similar existing methods, but also identify more than one type of leukocyte in a single blood smear image that contains multiple leukocyte types. To the best of our knowledge, this has been done for the first time and unlike other competitive methods, it does not limit the classification processes to only recognize a single leukocyte type in any given blood sample image.