Abstract
38
Introduction: We’ve developed a multiplexing scheme that takes advantage of Prism-PET’s deterministic light sharing to enable up to 36-to-1 crystal-to-channel coupling (i.e., 4 readout pixels per channel when using 9-to-1 crystal-to-pixel coupling). Prism-PET is a cost-effective high resolution PET detector module that has been shown to achieve sub-2 mm depth of interaction (DOI) resolution and on the order of 1 mm spatial resolution. In this work, we present a multiplexing scheme that takes advantage of Prism-PET’s unique deterministic optical properties to reduce readout data with minimal impact on module performance.
Methods: We acquired PET-like flood data on 4-to-1 and 9-to-1 coupled Prism-PET modules with a 3 MBq Na-22 point source (1 mm active diameter). The Prism-PET modules consisted of arrays of LYSO crystals (16 x 16 array of 1.4 x 1.4 x 20 mm crystals for 4-to-1, 24 x 24 array of 0.9 x 0.9 x 20 mm crystals for 9-to-1) coupled to 3 x 3 mm silicon photomultiplier (SiPM) pixels on one end and to a prismatoid light guide array on the opposite end. Signal multiplexing was performed via post-processing by integrating the readout signals across multiple pixels to achieve 16-to-1 crystal-to-channel multiplexing with 4-to-1 coupled Prism-PET and 36-to-1 multiplexing with 9-to-1 coupled Prism-PET (Figure 1). The multiplexing scheme was devised based on Prism-PET’s deterministic light sharing characteristics, which optically isolate intercrystal light sharing to a subset of nearest neighbor crystals. Thus, with the proposed multiplexing scheme shown in Figure 1, none of the signal from intercrystal light sharing will be lost since all multiplexed pixels are spaced at least 1 apart in each direction. We also acquired flood data at 19 different depths (1-19 mm in steps of 1 mm across the 20 mm length) with the 4-to-1 coupled Prism-PET module using lead collimation with a 1 mm pinhole to determine the DOI performance. A convolutional neural network (CNN) with a U-Net architecture was used for 8 x 8 flood histogram generation from a 16 x 1 input layer of multiplexed data (Figure 2). CNN training was carried out using 80% of the dataset. Adadelta was used for training optimization and mean-squared error was used as the loss metric. A batch size of 500 and 50 epochs were used for training. 20% of the dataset was used for model testing. Energy weighted average method was used to perform DOI localization.
Results: Following CNN training convergence, excellent crystal identification was observed at the center, edges and corners in the plotted flood histograms using the 16 x 1 multiplexed data for both the 4-to-1 and 9-to-1 coupled Prism-PET modules (Figures 3 and 4). Average DOI resolution across all measured depths was 2.32 mm full-width at half-maximum (FWHM) for non-multiplexed data and 2.73 mm FWHM for multiplexed data (Figure 5), representing only minor performance degradation. Conclusion: We’ve presented a smart, low-capacitance multiplexing scheme for Prism-PET and experimentally validated its efficacy. Our proposed multiplexing scheme further improves Prism-PET’s data and power efficiency, enabling cost-effective high resolution TOF-DOI readout. Figure 1. 2D schematic of anode (X) connections in the proposed multiplexing scheme. Figure 2. Schematic representation of a CNN (U-Net) used to generate an 8 x 8 flood map from 16 x 1 multiplexed data. Figure 3. Experimental flood histogram results of 4-to-1 coupled Prism-PET using the proposed multiplexing scheme with a CNN after 1 training epoch (left) and after training for 50 epochs (right) on 16 x 1 multiplexed data. Figure 4. Experimental flood histogram results of 9-to-1 coupled Prism-PET using the proposed multiplexing scheme with a CNN after 1 training epoch (left) and after training for 50 epochs (right) on 16 x 1 multiplexed data. Figure 5. Experimental results for depth of interaction resolution of a center crystal in a 4-to-1 coupled Prism-PET module at different depths (2-18 mm in steps of 4 mm) with and without multiplexed readout.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.