Comparing Deep Learning Performance for Chronic Lymphocytic Leukaemia Cell Segmentation in Brightfield Microscopy...

Area of Science:

• Biomedical Imaging
• Computational Biology
• Cell Biology

Background:

• Accurate cell detection in low-contrast brightfield microscopy images is challenging.
• Automatic cell detection is vital for quantitative cell morphology and migration studies.
• State-of-the-art image segmentation methods are required for precise cell boundary detection.

Purpose of the Study:

• To evaluate and compare eight advanced neural network architectures for segmenting cells in low-contrast brightfield microscopy images.
• To identify the optimal deep learning model for detecting and segmenting chronic lymphocytic leukemia (CLL) cells.
• To assess the impact of different segmentation methods on the analysis of cellular morphological features.

Main Methods:

• Comparison of eight neural network architectures (U-net, U-net++, PAN, MAN, LinkNet, FPN, DeepLabV3, DeepLabV3+) for image segmentation.
• Training networks for 1000 epochs using PyTorch and PyTorch Lightning.
• Utilizing watershed algorithm and three-class semantic segmentation for instance segmentation.
• Employing StarDist, a deep learning tool for object detection.

Main Results:

• U-net++ architecture with a ResNeSt-269 backbone achieved the highest Intersection over Union (IoU) score of 0.8902 for semantic segmentation.
• Statistically significant differences (p < 0.0001) were observed in mean cell characteristics (area, circularity, solidity, perimeter, radius, shape index) across different segmentation methods.
• While algorithms showed overall agreement with ground truth, distinct methods prioritized different morphological features.

Conclusions:

• The optimal choice of cell segmentation method depends on the specific application and the cellular traits under investigation.
• U-net++ demonstrates superior performance for semantic segmentation of CLL cells in this context.
• Different segmentation approaches can influence the quantitative analysis of cell morphology, necessitating careful method selection.