Local Label Point Correction for Edge Detection of Overlapping Cervical Cells

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Supervised deep learning relies heavily on accurate data labeling, which is challenging and prone to errors in manual annotation.
• Existing datasets often suffer from labeling inaccuracies, hindering the performance of deep learning models in tasks like edge detection and image segmentation.

Purpose of the Study:

• To introduce a novel method, Local Label Point Correction (LLPC), for improving the quality of annotations in image datasets.
• To enhance the accuracy of edge detection and image segmentation, particularly for complex structures like overlapping cells.

Main Methods:

• The LLPC method employs a three-step process: gradient-guided point correction, point interpolation, and local point smoothing.
• Object contour labels are refined by aligning annotated points with pixel gradient peaks.
• A local linear fitting approach is utilized for smoothing corrected edges to mitigate noise-induced irregularities.

Main Results:

• The LLPC method demonstrated an average precision improvement of 30-40% across multiple deep learning networks.
• A new, high-precision dataset of overlapping cervical cells (CCEDD) was created using the LLPC method for validation.
• Experiments confirmed that LLPC effectively enhances manual label quality and boosts the accuracy of overlapping cell edge detection.

Conclusions:

• The LLPC method offers a robust solution for correcting labeling errors in image datasets.
• This approach significantly improves performance in edge detection and image segmentation tasks, especially with challenging data.
• The developed method and dataset are expected to advance research in automated image annotation and analysis.