Few-shot adaptation for morphology-independent cell instance segmentation

TL;DR

This work tackles the problem of distribution shift in cell instance segmentation caused by diverse microscopy modalities and non-convex bacterial morphologies. It introduces Adaptive Omnipose, a few-shot supervised domain adaptation framework built on a segmentation strategy that predicts a boundary distance field $\phi$, gradient flow $\bm{u}$, and boundary score $z$, enabling rapid adaptation from a single to five annotated cells. The method adds contrastive distance loss $\mathcal{L}^{CD}$ and contrastive boundary loss $\mathcal{L}^{CB}$ to align target features with a $\mathcal{L}^{IS}$ base loss, achieving fast adaptation (about 3 minutes on an RTX 3090) and improved AP at IoU 0.5 across BP, BF, and Worm datasets, outperforming prior approaches like CellTranspose and Omnipose-FT. The approach demonstrates that morphology-independent segmentation can be achieved with minimal manual labeling and computational cost, making it practical for large and diverse microscopy collections.

Abstract

Microscopy data collections are becoming larger and more frequent. Accurate and precise quantitative analysis tools like cell instance segmentation are necessary to benefit from them. This is challenging due to the variability in the data, which requires retraining the segmentation model to maintain high accuracy on new collections. This is needed especially for segmenting cells with elongated and non-convex morphology like bacteria. We propose to reduce the amount of annotation and computing power needed for retraining the model by introducing a few-shot domain adaptation approach that requires annotating only one to five cells of the new data to process and that quickly adapts the model to maintain high accuracy. Our results show a significant boost in accuracy after adaptation to very challenging bacteria datasets.

Figures (2)

• Figure 1: Architecture approach. Illustration of our contrastive learning-based few-shot cellular instance segmentation approach.
• Figure 2: Qualitative results. Samples from the Worm (left two columns) and Bacteria Fluorescence (right two columns) datasets highlighting the variability of cell morphologies, and the corresponding segmentations by different approaches. GT indicates ground-truth.