Robust vessel segmentation in laser speckle contrast images based on semi-weakly supervised learning

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Laser Speckle Contrast Imaging (LSCI) is crucial for visualizing blood flow but faces challenges in vessel segmentation due to low signal-to-noise ratio and irregular vascular aberrations.
• Accurate vessel segmentation is vital for diagnosing and monitoring diseases affecting vasculature.

Purpose of the Study:

• To develop a robust semi-weakly supervised learning strategy for precise vessel segmentation in LSCI.
• To address limitations of existing methods in handling low signal-to-noise, small vessel sizes, and diseased vascular regions.

Main Methods:

• Utilized a semi-weakly supervised approach with manually labeled normal vessels and pseudo-labeled abnormal vessels.
• Employed DeepLabv3+ for segmentation, with continuously updated pseudo-labels during training.
• Incorporated a style translation network to test robustness against abnormal vessel-like noise.

Main Results:

• Achieved high performance with an Intersection over Union (IOU) of 0.8671 and Dice coefficient of 0.9288 in objective evaluations.
• Demonstrated superior performance in segmenting main and tiny vessels and maintaining blood vessel connectivity in subjective evaluations.
• Showcased significant robustness when noise mimicking abnormal vessels was introduced.

Conclusions:

• The developed semi-weakly supervised learning strategy offers high efficiency and robustness for LSCI vessel segmentation.
• This method presents a promising tool for assessing vascular morphological and structural features in clinical settings.