Abstract

Cell counting in microscopy images is vital in medicine and biology butextremely tedious and time-consuming to perform manually. While automatedmethods have advanced in recent years, state-of-the-art approaches tend toincreasingly complex model designs. In this paper, we propose a conceptuallysimple yet effective decoupled learning scheme for automated cell counting,consisting of separate counter and localizer networks. In contrast to jointlylearning counting and density map estimation, we show that decoupling theseobjectives surprisingly improves results. The counter operates on intermediatefeature maps rather than pixel space to leverage global context and producecount estimates, while also generating coarse density maps. The localizer thenreconstructs high-resolution density maps that precisely localize individualcells, conditional on the original images and coarse density maps from thecounter. Besides, to boost counting accuracy, we further introduce a globalmessage passing module to integrate cross-region patterns. Extensiveexperiments on four datasets demonstrate that our approach, despite itssimplicity, challenges common practice and achieves state-of-the-artperformance by significant margins. Our key insight is that decoupled learningalleviates the need to learn counting on high-resolution density maps directly,allowing the model to focus on global features critical for accurate estimates.Code is available at https://github.com/MedAITech/DCL.