SubtlePET

SubtlePET is an image processing software that employs a deep neural network-based algorithm, specifically a convolutional neural network with residual learning, to reduce noise and improve image quality in PET scans for various radiotracers. Its methodology involves predicting and separating noise from structural components, integrating into DICOM workflows to provide enhanced diagnostic images for clinical use by radiologists and nuclear medicine physicians.

SubtlePET is an image processing software, version V 2.0.0, designed for radiologists and nuclear medicine physicians. Its primary function is to enhance the quality of Positron Emission Tomography (PET) images by reducing noise. The software integrates into existing radiology workflows as a DICOM node, receiving DICOM 3.0 digital medical image data, processing it, and then forwarding the enhanced study to a selected destination, typically a PACS system or workstation. SubtlePET supports various radiotracers, including fluorodeoxyglucose (FDG), amyloid, 18F-DOPA, 18F-DCFPyL, Ga-68 Dotatate, and Ga-68 PSMA, for PET imaging. The core technology of SubtlePET utilizes a deep neural network-based algorithm, specifically a convolutional neural network employing a residual learning approach. This method analyzes each pixel's neighborhood to predict and separate noise components from structural components, thereby enhancing image structure while reducing noise. Performance testing, including software verification, validation, and a noise reduction bench test using representative human data, demonstrated a significant average increase in quantitative metrics, confirming its efficacy in noise reduction. The device has been updated to support additional tracers with its main PET machine learning model, while previous models specific to PET/CT and PET/MRI have been removed.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. What specific quantitative metrics were used to evaluate the 'significant average increase' in image quality, and what were the benchmarks for these metrics?; How does the performance of the updated single PET deep learning model compare to the previous separate models for PET, PET/CT, and PET/MRI in terms of noise reduction efficacy for the supported tracers?; Are there any known limitations or specific types of PET image artifacts that the SubtlePET algorithm is less effective at reducing or might potentially exacerbate?; What is the typical processing time added to the workflow per PET study when using SubtlePET, and how does this vary with image matrix size or complexity?; How has the training dataset for the deep neural network been curated to ensure robustness and generalizability across different PET scanner manufacturers and imaging protocols for the newly added radiotracers (18F-DOPA, 18F-DCFPyL, Ga-68 Dotatate, Ga-68 PSMA)?