Robust Vascular Segmentation for Raw Complex Images of Laser Speckle Contrast Based on Weakly Supervised Learning

Area of Science:

• Biomedical Imaging
• Medical Artificial Intelligence
• Microcirculation Research

Background:

• Laser speckle contrast imaging (LSCI) offers non-invasive, real-time analysis of local blood flow microcirculation with high resolution.
• Vascular segmentation in LSCI images is challenging due to noise, complex structures, and irregular aberrations in diseased areas.
• Limited annotated LSCI data hinders supervised deep learning applications for vascular segmentation.

Purpose of the Study:

• To develop a robust weakly supervised learning method for LSCI vascular segmentation.
• To design an effective deep neural network (FURNet) for accurate blood vessel segmentation.
• To address the difficulties in LSCI image annotation and improve deep learning applications.

Main Methods:

• Proposed a weakly supervised learning approach using threshold combinations and processing flows to generate ground truth data, avoiding manual annotation.
• Designed FURNet, a deep neural network integrating UNet++ and ResNeXt architectures.
• Trained and validated the FURNet model on constructed and unknown LSCI datasets.

Main Results:

• Achieved high-quality vascular segmentation with the trained FURNet model.
• Demonstrated good generalization capabilities across multiple LSCI datasets and scenarios.
• Successfully validated the method in vivo on tumor microcirculation before and after embolization treatment.

Conclusions:

• The developed weakly supervised method provides a novel approach for LSCI vascular segmentation.
• FURNet shows significant potential for accurate microcirculation analysis and AI-assisted disease diagnosis.
• This work advances the application of artificial intelligence in medical imaging and diagnostics.