Overlap-Aware Segmentation of ImageS (OASIS) is a new deep learning framework for separating overlapping objects in scientific images while preserving object-specific intensities. OASIS is the first such framework designed to specifically target regions of overlap through region-specific weights in its loss function. We show that appropriate choices of region and object-specific weights in OASIS's loss function can significantly improve both topological and intensity reconstruction in regions of overlap, even in cases where a faint object overlaps with a much brighter object. The result is improved pixel-level separation on an object-by-object basis, enabling accurate reconstruction of topological observables on objects that were originally dominated by overlap.

The initial application of OASIS is to the MIGDAL experiment, where faint electron recoil tracks must be disentangled from overlapping nuclear recoils in optical time projection chamber data. OASIS is interdisciplinary and has potential applications in astronomy (e.g. galaxy deblending) and in biomedical imaging.

A detailed paper describing OASIS and its performance is currently in preparation. Check back soon for updates and open-source software.