Machine-learning-based source localization for an intraoperative forceps-type positron emission counter

Area of Science:

• Medical Imaging
• Surgical Technology
• Machine Learning in Medicine

Background:

• Accurate intraoperative identification of metastatic lymph nodes in esophageal cancer surgery is crucial for selective lymphadenectomy.
• Current forceps-type positron emission counters (PECs) lack positional feedback, hindering precise quantification of 18F-FDG uptake in lymph nodes.

Purpose of the Study:

• To develop a position-sensitive detector and machine learning algorithm for accurate intraoperative source localization within a forceps-type PEC.
• To improve the quantitative accuracy and reliability of PEC systems for lymph node assessment during surgery.

Main Methods:

• A position-sensitive detector was created by segmenting a scintillator into a 2x2 crystal array.
• A machine learning model was trained using Monte Carlo simulations to estimate the 3D center of gravity (CoG) of a radioactive source.
• The system utilizes coincidence count values from detector pairs as input for the machine learning model.

Main Results:

• The machine learning model estimated the source CoG with a mean absolute error (MAE) of approximately 0.74 mm in simulations.
• Repositioning based on the estimated CoG reduced measurement variability from 52% to 15% in simulations.
• A prototype experiment achieved a Euclidean MAE of 1.33 mm.

Conclusions:

• Machine learning-based source localization significantly enhances the accuracy and reliability of forceps-type PEC systems.
• This technology holds substantial potential for improving intraoperative lymph node assessment in esophageal cancer surgery.
• The developed system offers improved quantitative accuracy for real-time surgical guidance.