Deep Learning Signal Discrimination for Improved Sensitivity at High Specificity for CMOS Intraoperative Probes.

Area of Science:

• Medical Physics
• Nuclear Medicine
• Artificial Intelligence

Background:

• Accurate delineation of cancerous tissue during surgery is critical for effective cancer resection.
• Intraoperative probes using radiotracers can aid in detecting cancer cells, but distinguishing signals from background noise is challenging.

Purpose of the Study:

• To explore deep learning algorithms for background gamma ray signal rejection in intraoperative probes.
• To evaluate the performance of convolutional neural networks (CNNs) for beta-gamma discrimination using CMOS monolithic active pixel sensors.

Main Methods:

• Two CNN-based methods were investigated: event cluster classification and semantic segmentation of event clusters.
• The methods were tested using radionuclides including Carbon-14 (¹⁴C), Cobalt-57 (⁵⁷Co), and Technetium-99m ([Formula: see text]Tc).
• Performance was evaluated using metrics such as the area under the curve (AUC) of the receiver operating characteristic (ROC) and the dice score.

Main Results:

• The classification deep network achieved an improved AUC of 0.93 for ¹⁴C beta and [Formula: see text]Tc gamma clusters, surpassing the conventional feature-based discriminator (0.88).
• A customized AUC loss function improved sensitivity by 31% at low false positive rates.
• The segmentation deep network achieved a mean dice score of 0.93.

Conclusions:

• Deep learning, particularly the classification method, shows superior performance for beta-gamma discrimination in intraoperative probes.
• These findings suggest that deep learning can significantly enhance the accuracy of intraoperative cancer detection, aiding surgeons in precise tissue excision.