Blood cell classification using interpretable shape features: A Comparative study of SVM models and CNN-Based approaches

Abstract

Background and Objective: Models to count blood cells have been extensively studied, but there are very few for classifying potential blood cells. This is crucial for verifying that what these systems are counting are in fact the correct object of interest. Furthermore, we have very little insight into which features are important for differentiating these classes in computational models.Methods:This paper presents a competitive solution for classifying red blood cells, white blood cells, and platelets from each other.Results:We were able to achieve an overall classification rate of about 99% and outperformed convolutional neural networks by about 10% by using the Blood Cell Count Dataset with a support vector machine, a polynomial kernel, and interpretable and explainable features. Further, we are also able to identify that the important features for classifying the three classes are the number of corners, the proportion of the area of the cell over the relevant area the cell occupies, the major axis relative length of the cell, and the relevant area surrounding the cell.Conclusion: We built a model for blood cell classification that is more interpretable and explainable while also outperforming the state-of-the-art.

Keywords

• Blood cell classification
• Shape classification
• Explainable machine learning