Sparse Regression Algorithm for Activity Estimation in $\gamma$ Spectrometry

Yann Sepulcre, Tom Trigano, Ya’acov RitovIEEE TSP 2013 · DOI · arXiv

Abstract

We consider the counting rate estimation of an unknown radioactive source, which emits photons at times modeled by an homogeneous Poisson process. A spectrometer converts the energy of incoming photons into electrical pulses, whose number provides a rough estimate of the intensity of the Poisson process. When the activity of the source is high, a physical phenomenon known as pileup effect distorts direct measurements, resulting in a significant bias to the standard estimators of the source activities used so far in the field. We show in this paper that the problem of counting rate estimation can be interpreted as a sparse regression problem. We suggest a post-processed, non-negative, version of the Least Absolute Shrinkage and Selection Operator (LASSO) to estimate the photon arrival times. The main difficulty in this problem is that no theoretical conditions can guarantee consistency in sparsity of LASSO, because the dictionary is not ideal and the signal is sampled. We therefore derive theoretical conditions and bounds which illustrate that the proposed method can none the less provide a good, close to the best attainable, estimate of the counting rate activity. The good performances of the proposed approach are studied on simulations and real datasets.

Introduction: The Challenge of Pulse Pileup

In nuclear spectroscopy, detectors convert incoming photons into electrical pulses. Measuring the "activity" of a radioactive source involves counting these pulses, which are generated by a random Poisson process. However, at high activity levels, photons arrive so rapidly that their corresponding electrical pulses overlap in time—a phenomenon known as pulse pileup.

Illustration of pulse pileup where overlapping electrical pulses (red section) distort individual photon signals.

Standard counting methods, such as simple thresholding or looking for "idle" periods where the signal returns to zero, fail under these conditions because they cannot distinguish individual pulses within a distorted signal "blob." This leads to a significant downward bias in activity estimation. While offline correction methods exist, they are often computationally intensive or rely on perfectly known pulse shapes, which are rarely available in real-world sampled digital signals.

Methodology: Activity Estimation as Sparse Regression

This paper reinterprets activity estimation as a sparse regression problem. Given that photons are relatively rare compared to high-speed electronic sampling rates, the recorded signal can be modeled as a sparse combination of pulse shapes at various time shifts.

Non-Negative LASSO (NNLASSO)

The authors employ a non-negative version of the LASSO (Least Absolute Shrinkage and Selection Operator). Non-negativity is a critical physical constraint, as photon energies and the resulting electrical pulses are inherently positive.

Dictionary and Post-Processing

A dictionary of "truncated gamma shapes" is constructed to model detector pulses. Since true arrival times rarely align perfectly with the discrete sampling grid, the dictionary is inherently "incomplete." To address this, the authors introduce a two-step post-processing logic:

  1. Thresholding: Removing negligible weights that likely represent noise fitting.
  2. Merging: Grouping consecutive active time blocks into a single "event." This allows the algorithm to correctly count one photon even if its energy is spread across adjacent sampling bins.

Theoretical bounds are derived to prove that even with an imperfect dictionary, the estimator remains consistent and stays close to the "best attainable" estimate.

Experimental Validation and Results

The approach was validated using both synthetic simulations and real spectrometric data from the ADONIS system.

Simulation Performance

Simulations tested Poisson processes with intensities ($\lambda$) from 0.1 to 0.4, representing very high physical activities. Two scenarios were evaluated: an "ideal" case where the signal matched the dictionary, and a "robustness" case with randomly varied pulse shapes.

Estimated versus actual values of lambda showing strong linearity and consistency across different activity levels.

As shown in the results, the estimated activity ($\hat{\lambda}$) maintains a near-perfect linear relationship with the actual values, demonstrating the method's reliability even at high counting rates. Compared to standard thresholding, the proposed algorithm significantly reduces variance.

Real-World Accuracy

On real data from the ADONIS system (sampled at 10 MHz), the algorithm successfully identified individual pulses within complex pileup events. For a source with a known activity, the algorithm yielded an estimate ($3.99 \times 10^4$ for the specific sample) that closely matched the benchmark, outperforming standard instruments.

Conclusion

By shifting the focus from signal reconstruction to coefficient counting, this sparse regression approach effectively overcomes the pileup effect in $\gamma$ spectrometry. While limited by the sampling rate (it cannot distinguish sub-sample arrivals), the method provides a robust, theoretically grounded framework for high-activity measurements. Future work will explore joint energy estimation to identify specific radioactive isotopes and adapt the model for non-homogeneous sources.

Summary updated

Abstract

We consider the counting rate estimation of an unknown radioactive source, which emits photons at times modeled by an homogeneous Poisson process. A spectrometer converts the energy of incoming photons into electrical pulses, whose number provides a rough estimate of the intensity of the Poisson process. When the activity of the source is high, a physical phenomenon known as pileup effect distorts direct measurements, resulting in a significant bias to the standard estimators of the source activities used so far in the field. We show in this paper that the problem of counting rate estimation can be interpreted as a sparse regression problem. We suggest a post-processed, non-negative, version of the Least Absolute Shrinkage and Selection Operator (LASSO) to estimate the photon arrival times. The main difficulty in this problem is that no theoretical conditions can guarantee consistency in sparsity of LASSO, because the dictionary is not ideal and the signal is sampled. We therefore derive theoretical conditions and bounds which illustrate that the proposed method can none the less provide a good, close to the best attainable, estimate of the counting rate activity. The good performances of the proposed approach are studied on simulations and real datasets.
BibTeX
@article{sepulcre2013sparse,
  author = {Yann Sepulcre and Tom Trigano and Ya’acov Ritov},
  title = {{Sparse Regression Algorithm for Activity Estimation in $\textbackslash\{\}gamma$ Spectrometry}},
  journal = {IEEE Transactions on Signal Processing},
  year = {2013},
  doi = {10.1109/tsp.2013.2264811},
}
Figure