Comparative Analysis of Deep Learning-Based Models for Blood Cell Image Segmentation

The accurate segmentation of blood cells is crucial in determining a wide range of hematological disorders. Therefore, in this paper, we investigated the most accurate and efficient blood cell segmentation approach by considering four state-of-the-art deep learning segmentation models, namely, Feature Pyramid Network (FPN), LinkNet, PSPNet, and UNet. The performance metrics used are IoU, Dice Coefficient, Precision, Recall, and F1-score, besides computation measures: execution time and memory usage. It is evident from the result that UNet was the best model in the category, by achieving an IoU of 0.85 and a Dice Coefficient of 0.92, thus guaranteeing superior quality of segmentation with a suitable execution time of 120ms. This paper presents a systematic analysis of segmentation quality along with the computation efficiency factor, and deep insights toward real-world medical image processing applications.