FDNet: Frequency Domain Denoising Network For Cell Segmentation in Astrocytes Derived From Induced Pluripotent Stem Cells

TL;DR

This work tackles the challenging task of astrocyte segmentation in Incucyte images where background interference obscures cells. It introduces the IAI704 dataset and a Frequency Domain Denoising Network (FDNet) that combines Contextual Information Fusion (CIF), an Attention Block (AB), and a Fourier Transform Block (FTB) to fuse multi-scale context and suppress interference in the frequency domain, using a high-pass filter with threshold $\eta = 0.3$. Evaluations on IAI704 show that FDNet outperforms seven state-of-the-art baselines with $mIoU = 80.88$ and Dice = 86.2, validating the method's effectiveness for astrocyte delineation during iPSC differentiation. The ablation studies corroborate the contributions of CIF, AB, and FTB, and the authors suggest that this approach can extend to other cell types and provide insights into differentiation progress relevant to neurodegenerative disease research, aided by the publicly released IAI704 dataset.

Abstract

Artificially generated induced pluripotent stem cells (iPSCs) from somatic cells play an important role for disease modeling and drug screening of neurodegenerative diseases. Astrocytes differentiated from iPSCs are important targets to investigate neuronal metabolism. The astrocyte differentiation progress can be monitored through the variations of morphology observed from microscopy images at different differentiation stages, then determined by molecular biology techniques upon maturation. However, the astrocytes usually ``perfectly'' blend into the background and some of them are covered by interference information (i.e., dead cells, media sediments, and cell debris), which makes astrocytes difficult to observe. Due to the lack of annotated datasets, the existing state-of-the-art deep learning approaches cannot be used to address this issue. In this paper, we introduce a new task named astrocyte segmentation with a novel dataset, called IAI704, which contains 704 images and their corresponding pixel-level annotation masks. Moreover, a novel frequency domain denoising network, named FDNet, is proposed for astrocyte segmentation. In detail, our FDNet consists of a contextual information fusion module (CIF), an attention block (AB), and a Fourier transform block (FTB). CIF and AB fuse multi-scale feature embeddings to localize the astrocytes. FTB transforms feature embeddings into the frequency domain and conducts a high-pass filter to eliminate interference information. Experimental results demonstrate the superiority of our proposed FDNet over the state-of-the-art substitutes in astrocyte segmentation, shedding insights for iPSC differentiation progress prediction.

Figures (3)

• Figure 1: Examples of our IAI704. From top to bottom are original Incucyte images and manually annotated labels. Some examples of interference (i.e., dead cells, media sediments and cell debris) are marked by red circles.
• Figure 2: Architecture of our proposed FDNet. Our model contains a backbone, a Contextual Information Fusion module CIF, an Attention Block AB and a Fourier Transform Block FTB.
• Figure 3: Visualization of astrocyte segmentation results with (a) as the input of the network. (b) is a ground-truth (GT) segmentation mask and (c-f) represent the segmentation masks generated by our proposed network, UCtransNet, UNet++ and UNet, respectively.