DeepSegNet: An Innovative Framework for Accurate Blood Cell Image Segmentation

Abstract

Image segmentation plays a crucial and indispensable role in computer vision, as it allows the partitioning of an image into meaningful regions or objects. Among its numerous applications, image segmentation holds particular significance in the domains of medical diagnosis and healthcare. Its vital role in this field stems from its ability to extract and delineate specific anatomical structures, tumors, lesions, and other critical regions from medical images. In medical diagnosis, accurate and precise segmentation of organs and abnormalities is paramount for effective treatment planning, disease monitoring, and surgical interventions. Blood cell image segmentation is highly valuable for medical diagnosis and research, particularly in the domains of hematology and pathology. Precisely segmenting blood cells from microscopic images is essential, as it offers critical insights into various blood-related disorders and diseases. Although deep learning segmentation models have exhibited promising results in blood cell image segmentation, they suffer from several limitations. These drawbacks encompass scarce data availability, inefficient feature extraction, extended computation time, limited generalization to unseen data, challenges with variations, and artifacts. Consequently, these limitations can adversely impact the overall performance of the models. Blood cell image segmentation encounters persistent challenges due to factors like irregular cell shapes, which pose difficulties in boundary delineation, imperfect cell separation in smears, and low cell contrast, leading to visibility issues during segmentation. This research article introduces the innovative DeepSegNet framework, a powerful solution for precise blood cell image segmentation. The performance of widely-used segmentation models like PSPNet, FPN, and DeepLabv3+ is enhanced through the use of sophisticated preprocessing techniques, improving generalization capability, data diversity, and training stability. Additionally, the incorporation of diverse dilated convolutions and feature fusion further contributes to the improvement of these models. The Improved PSPNet, Improved FPN, Deep Lab V3, and Improved Deep Lab V3+ achieved 98.25%, 99.04%, 98.23%, and 99.31% accuracy, respectively, and the Improved Deep Lab V3+ model outperformed well and produced a Dice Coefficient of 99.32% and Precision of 99.38%. The proposed DeepSegNet framework improves overall performance with an increased accuracy of 8.91%, 3.72%, 17.73%, 22.83%, 7.96%, 9.61%, 17.36%, 6.22%, 13.32%, and 14.32% compared to the existing models. This framework, which can be applied to accurately identify and quantify different cell types from blood cell images, is instrumental in diagnosing a variety of hematological disorders and diseases.