2DiNTS: 2D Differentiable Neural Network Topology Search for Multi-Modal Cell Segmentation

Hussam Azzuni, Muhammad Hamza Sharif, Naveed Syed, Mohammad Yaqub, Min Xu

Keywords: Multi-modal, Segmentation, Differentiable architecture search

TL;DR: Differentiable Neural Network Topology Search for Multi-Modal Cell Segmentation

Abstract: Cell segmentation is a crucial step in various biological and biomedical applications. However, microscopic images could vary depending on the lighting conditions employed. Training a deep learning model to handle various modalities could prove quite difficult since it will be much harder to create a comprehensive model. Our proposed method tackles four different microscopic imaging modalities such as brightfield, fluorescent, phase-contrast, and differential interference contrast to create a flexible multi-modal cell segmentation model. Differentiable Neural Network Topology Search (DiNTS) helps in providing a search space to accommodate the different features between the four modalities, which resulted in an increase of 0.5\% F1 on the validation set.