White blood cell classification using custom deep neural network and visualizing features of the images using heatmaps.

Blood count is a key method for diagnosing diseases by analyzing blood cell images using advanced equipment. Traditionally, this process involves invasive techniques and is time-consuming and costly. Modern approaches leverage deep learning (DL) to streamline this process, making it faster and more cost-effective. Given the similarity among blood cell images, distinguishing various blood types by sight is challenging. To address this, we propose a Customized deep neural network (CDNN) to accurately classify different types of blood cells while avoiding overfitting and degradation. CDNN uses a unique DL architecture to classify white blood cells (WBCs). We validated this architecture using the Raabin WBC and BCCD datasets. The model is optimized through preprocessing techniques such as normalization and data augmentation. Simulations were conducted with a batch size of 64, utilizing the Adam optimizer over 50 epochs. Our model achieved a high accuracy of 97.97% on the Raabin dataset and 99.64% on the BCCD dataset, outperforming existing state-of-the-art models. To understand the model's classification process, we applied G-CAM and LIME. These results suggest that CDNN can be developed into clinically useful solutions for detecting WBCs in blood cell images, significantly improving diagnostic efficiency.