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

Intraoperative identification of metastatic lymph nodes in esophageal cancer surgery could enable more selective lymph-node dissection. A forceps-type positron emission counter (PEC)-a compact coincidence detector designed to intraoperatively quantify 18F-FDG uptake in individual lymph nodes through standard laparoscopic trocars-requires the radioactive source to be centered within its field of view for accurate quantification, yet current hardware provides no positional feedback. Approach. A position-sensitive detector was designed by segmenting the conventional monolithic scintillator into a 2×2 crystal array. A pair of such detectors provides 16 coincidence count values, which serve as input to a machine-learning model that outputs the three-dimensional center of gravity (CoG) of the source with an intrinsic uncertainty indicator. Training data were generated by Monte Carlo simulations using a sensitivity-map superposition method with random source distributions varying in size, shape, position, and activity concentration. Main results. The CoG was estimated with errors of approximately 0.34-0.42 mm per axis (Euclidean mean absolute error (MAE) 0.74 mm). In simulation, repositioning based on the estimated CoG reduced measurement variability (percent standard deviation) from 52% to 15%. A prototype experiment achieved Euclidean MAE of 1.33 mm at 100 coincidence counts. Significance. These results demonstrate that machine-learning-based source localization has substantial potential to enhance the quantitative accuracy and reliability of forceps-type PEC systems for intraoperative lymph node assessment.&#xD.