Difference-Deformable Convolution With Pseudo Scale Instance Map for Cell Localization Cell localization still faces two unresolved challenges: 1) the dramatic variations in cell morphology, coupled with the heterogeneous intensity distribution of lightly stained cells; 2) existing cell location maps lack scale information, resulting in insufficient supervision for point maps and inaccurate supervision for density maps. 1) To address the first challenges, we introduce a novel gradient-aware and shape-adaptive Difference-Deformable Convolution (DDConv), which enhances the model’s ro- bustness to color by leveraging gradient information while adaptively adjusting the shape of the convolutional ker- nel to tackle the substantial variability in cell morphology. 2) To overcome the issue of unreasonable location maps, we propose the Pseudo-Scale Instance (PSI) map, which can adaptively provide the corresponding scale information for each cell to realize accurate supervision. We analyze and evaluate DDConv and the PSI map in three challenging cell localization tasks. In comparison to existing methods, our proposed approach significantly enhances localization performance, setting a new benchmark for the cell localiza- tion task.

Cell localization is an important area of medical image analysis, which is dedicated to predicting the precise location of cells in an image. The existing localization paradigm is to predict the density map using a Convolutional Neural Networks (CNN) model based on vanilla convolution and then process the density map using a local maximum search strategy to obtain the cell location and number information. However, there are three main problems in this paradigm: 1) CNN models based on vanilla convolution have difficulty in handling large variations in cell color; 2) The density map is difficult to provide accurate cell location information and ideal gradient information, and the information loss is more obvious in dense regions; 3) The post-processing strategy of density maps is susceptible to background noise and mutual interference between negative and positive cells. To tackle the above issues, we have made a comprehensive update of the existing paradigm, which consists of three parts: 1) A multi-scale gradient aggregation module based on difference convolution to effectively mitigate the challenge of large variations in cell color; 2) A new exponential distance transform map that provides accurate cell location information along with ideal gradient information for model learning; 3) A post-processing strategy named cell center localization strategy based on location maps that can significantly improve the localization performance. Extensive experiments on multiple datasets show that our approach can substantially improve cell localization and counting performance, establishing a new baseline for the cell localization task, and thereby increasing the efficiency of computer-aided diagnosis.