Split and Expand: An inference-time improvement for Weakly Supervised Cell Instance Segmentation

We consider the problem of segmenting cell nuclei instances from Hematoxylin and Eosin (H&E) stains with dot annotations only. While most recent works focus on improving the segmentation quality, this is usually insufficient for instance segmentation of cell instances clustered together or with a small size. In this work, we propose a simple two-step post-processing procedure, Split and Expand, that directly improves the conversion of segmentation maps to instances. In the splitting step, we generate fine-grained cell instances from the segmentation map with the guidance of cell-center predictions. For the expansion step, we utilize Layer-wise Relevance Propagation (LRP) explanation results to add small cells that are not captured in the segmentation map. Although we additionally train an output head to predict cell-centers, the post-processing procedure itself is not explicitly trained and is executed at inference-time only. A feature re-weighting loss based on LRP is proposed to improve our method even further. We test our procedure on the MoNuSeg and TNBC datasets and show quantitatively and qualitatively that our proposed method improves object-level metrics substantially.