Phytoplankton Image Segmentation and Annotation Method Based on Microscopic Fluorescence.

Area of Science:

• Marine Biology
• Environmental Science
• Computer Vision

Background:

• Accurate microscopic phytoplankton segmentation is crucial for water quality assessment.
• Current computer vision methods face challenges with background impurities and extensive manual annotation.
• Phytoplankton fluorescence characteristics offer a potential avenue for improved segmentation.

Purpose of the Study:

• To develop an automated method for segmenting and annotating phytoplankton contours.
• To fuse fluorescence and bright field microscopic images for enhanced segmentation accuracy.
• To reduce the significant manual annotation workload in phytoplankton image analysis.

Main Methods:

• Utilized phytoplankton fluorescence under excitation light combined with bright field imaging.
• Applied morphological operations on fluorescence images for initial contour detection.
• Employed Active Contour model on bright field images for contour refinement.

Main Results:

• The proposed method achieved recall, precision, F1 score, and IOU of 85.3%, 84.5%, 84.7%, and 74.6% respectively, outperforming manual labeling in initial tests.
• When used to train Mask-RCNN, automatically annotated data yielded superior results (97.0% recall, 86.5% precision, 91.1% F1, 84.2% IOU) compared to manual annotations (95.3% recall, 86.1% precision, 90.3% F1, 82.8% IOU).
• Demonstrated accurate segmentation of seven different algae species.

Conclusions:

• The proposed image fusion and active contour method accurately segments microscopic phytoplankton.
• Automatically annotated contour data is as effective as manually annotated data for training deep learning models like Mask-RCNN.
• This approach substantially decreases the manual annotation effort required for phytoplankton image analysis.