Towards efficient machine-learning-based reduction of the cosmic-ray induced background in X-ray imaging detectors: increasing context awareness

Pixelated detectors aboard modern space X-ray telescopes offer excellent, simultaneous imaging and spectroscopic capabilities. However, in orbit measurements with these devices are hampered by the background induced by charged particles, primarily cosmic-ray protons, interacting with the spacecraft and detectors, which makes the task of studying faint, low surface brightness objects challenging. The traditional background reduction algorithms analyze the deposited signal as a separate, multiple-pixel island, ignoring contextual information present on the same readout frame. Machine learning models can effectively capture this context by recognizing spatial and energetic correlations between signals left by a primary cosmic ray and secondaries produced during interaction of the incident particle with the body of the detector. In this work we present a study of different approaches to maximize the context information available to the deep neural network models to further improve the background reduction capabilities.