3D position estimation using an artificial neural network for a continuous scintillator PET detector

Abstract

Continuous crystal based PET detectors have features of simple design, low cost, good energy resolution and high detection efficiency. Through single-end readout of scintillation light, direct three-dimensional (3D) position estimation could be another advantage that the continuous crystal detector would have. In this paper, we propose to use artificial neural networks to simultaneously estimate the plane coordinate and DOI coordinate of incident γ photons with detected scintillation light. Using our experimental setup with an '8 + 8' simplified signal readout scheme, the training data of perpendicular irradiation on the front surface and one side surface are obtained, and the plane (x, y) networks and DOI networks are trained and evaluated. The test results show that the artificial neural network for DOI estimation is as effective as for plane estimation. The performance of both estimators is presented by resolution and bias. Without bias correction, the resolution of the plane estimator is on average better than 2 mm and that of the DOI estimator is about 2 mm over the whole area of the detector. With bias correction, the resolution at the edge area for plane estimation or at the end of the block away from the readout PMT for DOI estimation becomes worse, as we expect. The comprehensive performance of the 3D positioning by a neural network is accessed by the experimental test data of oblique irradiations. To show the combined effect of the 3D positioning over the whole area of the detector, the 2D flood images of oblique irradiation are presented with and without bias correction.

Self-Organizing Map Neural Network-Based Depth-of-Interaction Determination for Continuous Crystal PET Detectors

“…The basic idea of the nearest neighbor algorithm is to match the light distribution of an unknown event with the reference events by calculating their Euclidean distances [see (1)]. The position of the event can be determined as the position of its nearest neighbor among the reference events.…”

“…Recent research work on this issue reporting a two-dimensional (2D) plane position resolution better than 2 mm [1]- [3] shows that the topic using continuous crystal detectors in the future design of high resolution PET scanners is interesting and promising. Like pixelated PET detectors, depth-of-interaction (DOI) information in continuous crystal detector is also necessary to avoid the performance degradation caused by parallax.…”

“…The challenge on the position estimation is to find an effective algorithm to represent the mapping relationship. The latest research mainly focuses on the methods of multilayer perceptron (MLP) neural network-based position estimation [1], [6], statistics-based positioning (SBP) scheme [2], and nearest neighbor ( -NN) method [3]. All these methods statistically extract the mapping mechanism from pre-acquired reference events, and then used it to estimate the interaction position for unknown events.…”

“…Although the scheme has intrinsic DOI correction ability, the reference data acquisition in any possible incident angles would make the scheme impractical. In our previous work [1], we demonstrated MLP neural network-based 3D estimation scheme, which uses the perpendicular irradiation on front surface to train a neural network for the plane position estimation and uses the side irradiation to train a neural network for the DOI estimation. However, the reference data on side surface is also impractical in some cases, such as for crystals with a large size or for detector modules which have been installed 0018-9499 © 2015 IEEE.…”

Design of a high speed online maximum likelihood positioning algorithm using FPGA

“…We have built a prototype continuous crystal PET detector module based on a single block LYSO crystal coupled with a multi-anode PMT H7546B, and developed a simple test setup aiming at evaluating the performance of positioning algorithms [4]. Previously, we have reported improved statistical based positioning method (ISBP) [5], which uses real distribution from a large number of experimental samples instead of the assumption of Gaussian distribution, reveals better performance in event positioning.…”