Improved U-net-based leukocyte segmentation method.

Area of Science:

• Medical Imaging
• Computational Biology
• Hematology

Background:

• Accurate segmentation of leukocytes (neutrophils, basophils, eosinophils, monocytes, lymphocytes) is crucial for disease diagnosis.
• Blood cell image acquisition is challenging due to environmental factors causing poor image quality and indistinct leukocyte features.

Purpose of the Study:

• To develop an improved U-net based method for robust leukocyte segmentation in complex, variable-quality images.
• To enhance feature clarity, handle inter-class similarity, and address class imbalance in blood cell image segmentation.

Main Methods:

• Employed adaptive histogram equalization-retinex correction for data enhancement.
• Integrated a convolutional block attention module into U-net to improve feature focus and network efficiency.
• Utilized a combined focal loss and Dice loss function to manage class imbalance and improve cytoplasm segmentation.

Main Results:

• Achieved 99.53% accuracy and 91.89% mIoU on the BCISC public dataset for segmenting multiple leukocytes.
• Demonstrated effective segmentation of lymphocytes, basophils, neutrophils, eosinophils, and monocytes.

Conclusions:

• The proposed improved U-net method significantly enhances leukocyte segmentation accuracy in challenging image conditions.
• The approach effectively addresses common issues in blood cell image analysis, offering a valuable tool for diagnostic applications.