A Power Efficient Solution to Determine Red Blood Cell Deformation Type Using Binarized DenseNet

Red Blood Cells (RBCs) play an important role in the welfare of human being as it helps to transport oxygen throughout the body. Different RBC-related diseases, for example, variants of anemias, can disrupt regular functionality and become life-threatening. Classification systems leveraging CNNs can be useful for automated diagnosis of RBC deformation, but the system can be quite resource-intensive in case the CNN architecture is large. The proposed approach provides an empirical analysis of the application of 28 and 45-layer Binarized DenseNet for identifying RBC deformations. According to our investigation, the accuracy of the 45-layer binarized variant can reach 93–94%, which is on par with the results of the conventional variant, which also achieves 93–94% accuracy. The 23-layer binarized variant, while not on par with the regular variant, also gets very close in terms of accuracy. Meanwhile, the 45-layer and 28-layer binarized variant only requires 9% and 11% storage space respectively to that of regular DenseNet, with potentially faster inference time. This optimized model can be useful since it can be easily deployed in resource-constrained devices, such as mobile phones and cheap embedded systems.