Speaker
Description
The MÖNCH detector [1], a charge integrating hybrid pixel detector with 25 um pixel pitch, has demonstrated micrometer-scale spatial resolution [2] via the η-interpolation method, enabling applications like energy-resolved X-ray imaging. However, this performance is restricted to low photon fluxes (~105 ph/s/mm2), as η-interpolation relies on spatially isolated single-photon events, hindering its use at high-brightness fourth-generation synchrotrons.
To overcome this limitation, we have developed a deep learning (DL)–based approach for sub-pixel localization of both single and pile-up X-ray events. Our DL models are trained on high-fidelity simulations that accurately model key physical processes including charge drift, diffusion, and Coulomb repulsion [3]. For the single-photon events, the DL model outperforms η-interpolation in localization accuracy by enabling absorption-depth–dependent charge-sharing profiles, rather than using a single, depth-averaged η distribution. More importantly, the DL model uniquely enables the reconstruction of two-photon pile-up events with sub-pixel resolution, thereby extending the MÖNCH’s usable flux range while preserving high spatial resolution.
We will present the detailed deep learning models and simulation-based training results quantitatively compared to the η-interpolation. We will also show experimental validation results, including an imaging phantom that directly compares the spatial resolution achieved with η-interpolation versus our DL-based method.
[1] M. Ramilli et al 2017 JINST 12 C01071
[2] S Cartier et al 2014 JINST 9 C05027
[3] X. Xie et al 2026 NIMA 1081 170894