Lidacs: a Lightweight Domain adaptive cell Segmentation Framework

Prakhar Rai, Shiv Gehlot, Ritu Gupta, Anubha Gupta

Abstract: This paper presents a LIghtweight Domain Adaptive Cell Segmentation (LIDACS) framework that achieves state-of-the-art results (0.9505 mIoU) in instance segmentation on the SegPC-21 challenge dataset featured in ISBI 2021, while being significantly parameter efficient than the existing methods. LIDACS is a hierarchical multi-stage approach that utilizes prior domain-specific information to perform statistical and empirical analysis. It also employs task-specific augmentations and improved transfer learning via shared representation to enable better data representation. LIDACS also applies a novel cell structure-based contrastive augmentation paired with cell cloning, increasing annotation density and promoting better stain color in-variance. Effectively, LIDACS is a lightweight architecture, efficient for practical deployment, that provides optimal generalization.