Cell image instance segmentation based on PolarMask using weak labels.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Accurate blood cell segmentation is crucial for medical diagnostics.
• Existing methods may struggle with contour smoothness and accuracy.
• Automated segmentation can reduce manual workload for healthcare professionals.

Purpose of the Study:

• To develop an improved PolarMask-based method for accurate blood cell contour segmentation.
• To enhance the speed and precision of blood cell segmentation models.
• To provide smoother and more accurate contours for erythrocytes and leukocytes.

Main Methods:

• Utilized a two-part approach: weak label-based pretraining and improved PolarMask segmentation.
• Employed bounding box labels for pretraining and segmentation labels for transfer learning.
• Incorporated a smoothing constraint loss and SE attention mechanism (ResNet18) for enhanced accuracy.

Main Results:

• Achieved significant improvements in average precision (AP) metrics compared to standard PolarMask.
• Increased AP by 8.4%, AP50 by 0.6%, and notably AP75 by 8.8%.
• Demonstrated superior performance in segmenting erythrocyte and leukocyte contours.

Conclusions:

• The proposed method effectively segments red and white blood cells using PolarMask with weak label pretraining.
• The technique significantly boosts segmentation accuracy and produces smoother cell contours.
• This advancement aids medical personnel in rapid cell identification and analysis.