Technical and functional design considerations for a real-world interpretable AI solution for NIR perfusion analysis...

Area of Science:

• Surgical technology
• Medical imaging
• Artificial intelligence in medicine

Background:

• Near-infrared (NIR) fluorescence imaging with indocyanine green (ICG) is clinically used for assessing tissue perfusion during surgery.
• Current applications primarily focus on tissue health checks before anastomoses.
• There is growing interest in utilizing fluorophores for intraoperative cancer detection, often relying on static imaging post-administration.

Purpose of the Study:

• To enhance the utility of NIR fluorescence assessment through interpretable artificial intelligence (AI) methods.
• To improve dynamic interpretation accuracy for tissue perfusion and enable new applications like real-time cancer detection.
• To develop a software as a medical device for in situ cancer characterization and broader tissue perfusion applications.

Main Methods:

• Generating fluorescence intensity curves from intraoperative NIR video streams.
• Processing curves to correct for image disturbances and extracting key features for tissue characterization.
• Utilizing machine learning classifiers trained on extracted features for tissue classification, including cancer detection.

Main Results:

• Demonstrated potential for AI to enable actionable predictions during the dynamic phase of ICG fluorescence (seconds to minutes post-administration).
• Research indicates AI methods can differentiate cancer from benign tissue in real-time based on differential fluorescence signaling.
• The interpretable methodology allows for accurate classification with modest training sets compared to deep learning.

Conclusions:

• Interpretable AI applied to dynamic NIR fluorescence imaging can significantly improve surgical assessment of tissue perfusion and enable real-time cancer detection.
• This approach offers a pathway for developing medical devices compliant with regulations, accessible to surgeons without specialized IT training.
• The developed methodology has broad relevance for in situ cancer characterization and other tissue perfusion applications in surgery.