Data-Centric Learning Framework for Real-Time Detection of Aiming Beam in Fluorescence Lifetime Imaging Guided...

Area of Science:

• Medical Imaging
• Surgical Technology
• Computer Vision

Background:

• Real-time surgical guidance is crucial for precision interventions.
• Fiber-based fluorescence lifetime imaging (FLIm) offers advanced tissue characterization.
• Accurate aiming beam detection is vital for mapping FLIm data in the surgical field, especially in challenging environments like Transoral Robotic Surgery (TORS).

Purpose of the Study:

• To develop and validate a robust method for detecting the aiming beam in FLIm for real-time surgical guidance.
• To address challenges posed by variable illumination and reflections in surgical settings.
• To improve the reliability and accuracy of image-guided interventions.

Main Methods:

• Developed an instance segmentation model using a data-centric training strategy.
• Focused on minimizing label noise and enhancing detection robustness for aiming beam identification.
• Evaluated the model on 40 in vivo surgical videos and during clinical TORS procedures.

Main Results:

• Achieved a median aiming beam detection rate of 85% on in vivo surgical videos.
• Maintained an 85% detection rate when the system was integrated into clinical TORS procedures.
• The system demonstrated computational efficiency of approximately 24 frames per second, suitable for real-time guidance.

Conclusions:

• The data-centric approach significantly enhances the reliability of FLIm-based aiming beam detection in complex surgical scenarios.
• This advancement supports the feasibility of real-time, image-guided interventions, improving surgical precision.
• The developed system is computationally efficient and effective in clinical settings like TORS.