Deep learning based semantic segmentation of leukemia effected white blood cell.

Medical image segmentation has numerous applications in diagnosing different diseases. Various types of diseases are found in white blood and Red blood cells. This paper represents the segmentation of WBCs from blood smear images. It is a complex and challenging task due to the frequent overlapping and variants in size and shape of WBCs with each other and RBCs. This overlapping is due to the rough border of the immature cells. The paper describes a new approach to WBC segmentation using UNet++, the marker watershed algorithm, and Neural Ordinary Differential Equations (ODE). This technique uses UNet++ for pre-segmentation, followed by the marker watershed method, which has been integrated using ODE to deepen the segmentation process. This novel integration enhances clinical applications in automated blood cell analysis, diagnostic imaging, and disease monitoring, improving accuracy and robustness. The ODE is used after the convolution operation to reduce the error at each step, preventing the massive propagation of error in the forward and the backpropagation. The White blood cells are segmented from the input smear images using ALL_IDB1 and ALL_IDB2 datasets, which are further used in the experiment section. UNet ++ is used to generate the pre-segmented probabilistic grayscale images. Some white blood cells are connected and make groups appearing in the grayscale images. These groups of WBCs are separated using a technique called the marker watershed, which gives us the final segmented result. The experimentation results show that the mean intersection over union (Jaccard method), the Dice similarity coefficient, and the mean pixel accuracy are 97.73%, 98.36%, and 98.97%, respectively. The structure and size of the white blood cells vary from red blood cells and platelets, which makes this work different from others. Furthermore, the combination of UNet++, marker watershed, and Neural Ordinary Differential Equation makes the proposed system unique from existing systems. This work can be further investigated to reduce computational complexity and memory space for optimizing deployment on low-resource devices, such as smart healthcare systems. Techniques like model pruning, quantization, or learned information distillation might be explored to create a lightweight version of the model without much loss in accuracy. Such developments would make possible mass uses of automated white blood cell segmentation in portable, low-cost health devices for point-of-care remote diagnostics and monitoring.