Deep-Learning based Blood Cells Classification and Initial Edge Device Implementation

The suitability of two deep learning models for classifying microscopic medical images is investigated on four types of white blood cell images; the inference phase is imple-mented onto a single board computer. The first proposed model is custom-developed by careful selection of the hyperparameters; the second one is based on transfer learning. A dataset of 12500 images is used. For the custom model, the training and validation accuracies are 99% and 97%, respectively, with overall classification accuracy of 97.77%. For the transferred model, the training and validation accuracies are 92% and 87%, respectively, with overall classification accuracy of 92% The two models are minimized and deployed on an RPi4 (quad-core ARM Cortex-A72 processor) for classification inference; the overall accuracy and classification times are 98.54% & 48.2 ms for the custom model vs. 91.3% & 142 ms for the transferred model, confirming that the custom model outperforms the transferred one.