Modified UNet++ with atrous spatial pyramid pooling for blood cell image segmentation

Area of Science:

• Medical imaging analysis
• Computational pathology
• Artificial intelligence in healthcare

Background:

• Accurate blood cell segmentation is vital for computer-aided diagnosis.
• Existing methods struggle with low contrast, varied cell morphology, and limited labeled data.
• Current segmentation performance is insufficient for clinical diagnostic requirements.

Purpose of the Study:

• To develop an advanced deep learning approach for improved blood cell segmentation in pathological images.
• To overcome limitations of low contrast, morphological variations, and data scarcity in cell segmentation.
• To enhance the accuracy and reliability of automated cell segmentation for diagnostic applications.

Main Methods:

• Utilized UNet++ as the backbone for multi-scale feature extraction.
• Redesigned skip connections to mitigate degradation and reduce computational complexity.
• Incorporated Atrous Spatial Pyramid Pooling (ASSP) for multi-scale feature acquisition.
• Implemented Multi-Sided Output Fusion (MSOF) to integrate features from diverse semantic levels.

Main Results:

• Achieved significant improvements in segmentation accuracy metrics, including Matthew's correlation coefficient (Mcc), Dice, and Jaccard values.
• Demonstrated superior performance compared to classical semantic segmentation networks on the BCISC dataset.
• The proposed method effectively handles challenges posed by low contrast and morphological variations in blood cell images.

Conclusions:

• The developed deep learning model offers a substantial advancement in blood cell image segmentation.
• The novel architecture and fusion strategy effectively address key limitations in pathological image analysis.
• This approach holds promise for improving the accuracy and efficiency of computer-aided diagnosis in hematology.