Enhancing Cell Instance Segmentation in Scanning Electron Microscopy Images via a Deep Contour Closing Operator

TL;DR

This work tackles the difficulty of accurately segmenting and distinguishing individual cells in scanning electron microscopy images, where incomplete cell contours impede automatic protein-free boundary delineation. It introduces a two-stage pipeline: Step 1 uses nnU-Net to generate voxelwise cell contour probability maps, and Step 2 applies a CNN-based closing operator (COp-Net) to fill gaps in those maps within an iterative refinement scheme. A novel PDE-based data augmentation strategy simulates realistic local degradations in contour probability, enabling robust training of the closing network. Across private PDX SEM datasets and public HeLa images, the COp-Net extension significantly improves correctly labelled cells and contour fidelity (NSD and clDice), while reducing manual correction time, demonstrating practical impact for quantitative tumor tissue analyses.

Abstract

Accurately segmenting and individualizing cells in SEM images is a highly promising technique for elucidating tissue architecture in oncology. While current AI-based methods are effective, errors persist, necessitating time-consuming manual corrections, particularly in areas where the quality of cell contours in the image is poor and requires gap filling. This study presents a novel AI-driven approach for refining cell boundary delineation to improve instance-based cell segmentation in SEM images, also reducing the necessity for residual manual correction. A CNN COp-Net is introduced to address gaps in cell contours, effectively filling in regions with deficient or absent information. The network takes as input cell contour probability maps with potentially inadequate or missing information and outputs corrected cell contour delineations. The lack of training data was addressed by generating low integrity probability maps using a tailored PDE. We showcase the efficacy of our approach in augmenting cell boundary precision using both private SEM images from PDX hepatoblastoma tissues and publicly accessible images datasets. The proposed cell contour closing operator exhibits a notable improvement in tested datasets, achieving respectively close to 50% (private data) and 10% (public data) increase in the accurately-delineated cell proportion compared to state-of-the-art methods. Additionally, the need for manual corrections was significantly reduced, therefore facilitating the overall digitalization process. Our results demonstrate a notable enhancement in the accuracy of cell instance segmentation, particularly in highly challenging regions where image quality compromises the integrity of cell boundaries, necessitating gap filling. Therefore, our work should ultimately facilitate the study of tumour tissue bioarchitecture in onconanotomy field.

Figures (8)

• Figure 1: Overview of the proposed pipeline for cell instance segmentation. A segmentation neural network first generates a probability map $u$ of cell contours from a 3D SEM stack input (Step #1, see section \ref{['cell contour segmentation']}). Subsequently, the proposed COp-Net automatically identifies and fills regions with insufficient or missing information (Step #2, see section \ref{['Cell contour closing operator']}). Finally, a connected component labelling algorithm is applied to achieve individual cell identification and produce the output cell instance segmentation.
• Figure 2: Schematic view of the proposed cell contour COp-Net. A 2D pixelwise continuous probability map $u$ obtained from Step #1 is taken as input. The proposed operator is composed by a closing network embedded in an iterative scheme.
• Figure 3: Typical data used for the end-to-end supervised training of the cell contour closing network. (a): private tumour SEM image, (b): corresponding ground truth cell contour , (c): simulated cell contour probability map derived from (b) using Eq. (\ref{['eq1']}), (d): zoom into the red square in (c) highlighting a global and local isotropic diffusion of cell contour probability, (e): zoom into the purple square in (c) highlighting a local lack of cell contour probability.
• Figure 4: Private dataset preparation and distribution used to train, validate and test the COp-Net. T1 and T2 represent two different PDX tumour tissue samples. The number of cell delineations is specified for both the training and testing datasets.
• Figure 5: Visual comparison of results obtained on the private testing dataset #2. (a) : private SEM image, (b-f): white colour represents cell contour, red colour represents blood capillary (BC) and haemorrhagic zone (HZ) obtained through manual segmentation, coloured cells represents accurately labelled cells compared to ground truth (b), and black indicates errors as compared to ground truth, (c): cell contours were obtained using the produced probability map by the Cellpose and a contour function.