Speaker
Description
Background. Organ-on-chip (OOC) platforms mimic human tissue at millimeter scale and provide unique models for drug development and disease research. Quantitative imaging of radiotracer distribution within these devices can totally change their scientific utility, however no PET system exists at the physical scale and resolution required. We present a dedicated on-chip PET detector concept. It revolves around a new dual-sided readout architecture, whereby SiPM arrays placed on one large face and on the side face of the same monolithic LYSO crystal yield complementary spatial information, drastically enhancing scintillation location determination and image reconstruction quality.
Methods. The physical prototype features two monolithic LYSO crystals (5×5 cm² in area, 15 mm thick) mounted face-to-face, with the OOC unit in the space between them. Dual-sided readout is carried out by each crystal: the large 5×5 cm² face on one side is equipped with a 16×16 array of 256 SiPM elements, while the 15 mm side face is covered by two 8×8 SiPM matrices, with channels not having direct crystal coverage being excluded from the analysis chain. Light patterns across the entire SiPM geometry were generated by GATE Monte Carlo simulations to train a Convolutional Neural Network (CNN) for predicting gamma-ray interaction points. The consistent part of large-face-only readout versus combined large-face and side-face readout was measured, thus directly quantifying the supplementary localization information encoded by orthogonal side-face matrices. Then, the simulation-trained CNN was used on experimental data from the physical prototype to confirm the generalization from simulated to real detector data. Spatial resolution was further restored by applying a deep learning positron-range correction algorithm based on U-Net architecture.
Results. The dual-sided readout design clearly brought a significant enhancement over the large-face-only layouts: side-face SiPM matrices provided depth-of-interaction information which could not be obtained from the large face only, thereby reducing the positioning error. In the optimized setup, the mean positioning error was 0.80 mm, the system sensitivity was 34.81%, and the mean spatial resolution of the reconstructed image was 0.55 mm FWHM according to the simulation. The positron-range correction using deep learning significantly enhanced spatial resolution by 32%, thus it recovered over 91% of the maximum theoretically achievable gain. Most importantly, the simulation-trained network when applied to the real experimental data still showed reliable positioning results which implies that the Monte Carlo training framework generalizes to the physical detector measurements without the need for retraining.
Conclusion. The implementation of a two-crystal face-to-face configuration with dual-sided readout a 16×16 large-face panel combined with 8×8 side-face arrays for each crystal is the main hardware revolution of this system. Light distributions recorded at the same time from perpendicular crystal faces represent spatial coordinates in a complementary way, thus sub-millimetre reconstruction can be achieved without pixelated detectors. Establishing the workability of the system, the transport of simulation-trained AI models to real experimental data has been attested as effective. This detector design, merged with AI-based reconstruction, is capable of yielding the quantitative and spatial results necessary for the insertion of functional PET imaging directly with organ-on-chip devices, which allows radiotracer-based measurement of metabolic and pharmacokinetic processes in living microfluidic tissue models.
| Track | FTMI |
|---|---|
| Presentation type | Oral |