A Benchmark Study of DeepLabV3+, U-Net++, and Attention U-Net for Blood Cell Segmentation

Abstrak

Cell segmentation is a critical process in biomedical image analysis. This study evaluated the performance of three state-of-the-art deep learning models—DeepLabV3+, U-Net++, and Attention U-Net—using the Blood Cell Count and Detection (BCCD) dataset, which contains annotated images of blood cells. The models were rigorously analyzed through qualitative and quantitative evaluations, employing accuracy, precision, recall, and F1 score metrics. The results demonstrated that all three models achieved high segmentation performance, with U-Net++ excelling in accuracy (0.9740), precision (0.9511), and F1 score (0.9576), Attention U-Net achieving the highest recall (0.9692), and DeepLabV3+ providing a balanced performance across all metrics. Qualitative analyses revealed that U-Net++ delivered superior segmentation of complex and overlapping cell structures, while Attention U-Net exhibited exceptional sensitivity to dense cell clusters. Training and validation curves of the models confirmed their stability and generalizability, indicating efficient convergence without overfitting. By highlighting the unique strengths of each model, this study emphasized the importance of selecting architectures tailored to specific tasks. Future research will expand the application of these models to diverse biomedical datasets to further advance automated image analysis and its impact on healthcare outcomes.