Attention-based Instance Segmentation Network for Cell Segmentation

Abstract

Segmentation of cell images has been widely explored in medical image analysis and clinical aided diagnoses. Instance segmentation technology, which can detect and distinguish each cell object has been a hot topic in recent years. However, the complex background interference and densely distribution in cell images always make the instance segmentation challenging. This study proposes a new instance segmentation algorithm based on the typical Mask R-CNN, named as attention-based instance segmentation network (AIS-Mask) for more accurate cell segmentation. Specially, an attention module is introduced and imposed on the top-down multi-scale information flow of the feature pyramid network (FPN) for extracting more efficient features and suppressing the background interference simultaneously. Experiments on the cell dataset from the Chinese cargo spacecraft TZ-1 demonstrate the impressing performance of our AIS-Mask under complex background interference. Both the quantitative evaluation and qualitative visual results show that the proposed AIS-Mask outperforms the state-of-the-art Mask R-CNN.