Performance Analysis of UNet Backbones with Dropout for Morphometric Measurement of Blood Cells

Measurement of Morphometric Parameters of Blood Cells (MPBC) plays an important role in effective hematological diagnosis in both humans and animals. Several deep learning models have been developed for segmenting blood cells to carry out MPBC measurements, especially UNet is emerging as a promising solution. However, with various UNet backbones implementations available, a comprehensive comparison of their performances remains unexplored. This paper deals with the performance evaluation of UNet backbones for measuring MPBC in the human white blood cell (WBC) dataset. Deep learning models including UNet are susceptible to overfitting, particularly when dealing with limited data. In this context, dropout plays a critical role, known for its effectiveness in reducing overfitting and boosting model generalization, thereby improving the model's overall performance. In the case of the WBC dataset, where data samples are typically limited, this work incorporated dropout in UNet backbones to evaluate the model's performance for cell area measurement and compare the performance across different backbones. Results show that EfficientNetB1, in comparison with the other backbones, has fewer outliers and lower variability in cell area measurements with an average standard deviation of 55 pixels. The average mean IoU score achieved across all UNet backbones is approximately 95.92%, demonstrating the correctness of spatial positioning of the predicted cell pixels. The main objective is to investigate the measurement reliability of these tools, which ultimately aid clinicians in making diagnostic decisions.