Development of a Compton Camera for Biological and Medical Imaging (SCoKa)

This research is dedicated to the fundamental challenge of detecting and analysing Cherenkov events in the Compton camera's scatter layer through the development
of robust, model-based algorithms. The primary objective is to achieve a high-fidelity reconstruction of the three-dimensional Cherenkov cone emitted by an energised Compton electron. This is accomplished by precisely analysing the two-dimensional Cherenkov ellipse projected onto the detector plane. The accurate determination of the cone's properties is paramount, as these parameters directly encode the direction and energy of the Compton electron, which are essential for the final reconstruction of the gamma-ray source.
In contrast to the learning-based methods also explored within this project, our approach adapts classical computer vision and data analysis techniques to the unique nature of the detector's output. The raw data is not a conventional image but a sparse, unstructured point cloud of individual photon hits, which requires tailoring algorithms to operate directly on coordinate-based data. This methodology is realised as a multi-stage processing pipeline, beginning with noise filtering and event clustering. This culminates in the geometric inversion process, where the ellipse's estimated parameters—its centre, size, and orientation—are used to mathematically back-project the shape into three-dimensional space and reconstruct the original Cherenkov cone.
Ultimately, the goal of this research is to develop a highly efficient and competitive solution capable of real-time event reconstruction. To achieve this, our algorithms are being co-designed with their hardware implementation in mind, specifically targeting deployment on dedicated hardware accelerators such as FPGAs (Field-Programmable Gate Arrays). This hardware-centric approach is crucial for harnessing the massive parallelism required to process event data with the ultra-low latency demanded by practical imaging applications. This synergy between an interpretable, model-based algorithm and an optimised hardware architecture is crucial for creating a robust, power-efficient system that can perform on-the-fly event reconstruction, positioning this pathway as a powerful solution for next-generation Compton cameras.

This research focuses on detecting Cherenkov events in the scatter layer of the Compton camera using machine learning approaches. When high-energy gamma rays interact within this layer, they generate Cherenkov photons that form circular or elliptical patterns on the detector. Accurately identifying these patterns is crucial for reconstructing the direction of the incoming gamma rays and determining their source position.
Traditional methods such as the Hough transform and ellipse fitting have been studied in earlier works, but they face limitations in handling noise and require extensive computation. To overcome these challenges, modern machine learning
architectures, including Convolutional Neural Networks (CNNs), are explored to learn and recognize these geometric patterns directly from detector data.
The models are trained and validated using both simulated and realistic datasets that represent detector responses under different conditions. This ensures robust performance and adaptability of the algorithms to experimental data.
In parallel, this research investigates the deployment of these models on hardware accelerators such as FPGAs for real-time and power-efficient processing. Using tools like FINN, the models are optimized through quantization and hardware-specific acceleration. This integration of AI and hardware design enables ultra-low-latency inference, bringing the Compton camera closer to real-time imaging applications in medical and biological diagnostics.

Image reconstruction focuses on various traditional reconstruction algorithms like Simple Back Projection (SBP), Filtered Back Projection (FBP), Maximum Likelihood Expectation Maximization (MLEM), Stochastic Origin Ensemble Method are explored. These algorithms are implemented starting from a point source reconstruction to increasing the complexity to 2D distribution and finally 3D distribution of the source.
The reconstruction algorithms work with the raw measurement data from the detector. Where, we get information of the co-ordinates for the point of intersection in both scatter plane P1 (sx,sy,sz) and absorption plane P2 (ax,ay,az). We also get information of scattering angle (β), scattered electron energy (Ee) and scattered gamma energy (Eγ). All these parameters measured from the detector have uncertainties included in the measurements due to the physical, geometrical and sensitivity limitations of the detector. These uncertainties must be studied and included in the reconstruction algorithms.
On the other hand, Neural Network models both machine learning and deep learning are explored to obtain direct reconstruction from the raw data. Multiple models like Random Forest, Transfer learning model, XGB boost were tested. But none of these models showed any good results for the required prediction. A couple of hybrid architectures with convolution layers, Dense layers, transformer and LSTM layers were tested and observed high accuracy results for ideal events. These architectures are further being improved for more realistic data with uncertainties.