Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Semi-Unsupervised Microscopy Segmentation with Fuzzy Logic and Spatial Statistics for Cross-Domain Analysis Using a GUI

Surajit Das, Pavel Zun

TL;DR

This work addresses the challenge of segmenting unstained live cells in low-contrast bright-field microscopy without annotations or training data. It introduces the Homogeneous Image Plane concept and a one-time calibration–driven unsupervised pipeline that combines SSDLM-based homogeneity measures, fuzzy inference, Moran’s I, and an adjusted variogram to separate background texture from cell morphology. The approach yields cross-domain robustness across bright-field, phase-contrast, and laser-trace data, achieving up to a 48% IoU gain over state-of-the-art methods while running on CPU and offering a GUI for non-programmers. These results establish a practical, training-free alternative for live-cell imaging analysis with significant potential for regenerative medicine and clinical microscopy workflows, while also outlining avenues for automation and broader modality validation.

Abstract

Brightfield microscopy of unstained live cells is challenging due to low contrast, dynamic morphology, uneven illumination, and lack of labels. Deep learning achieved SOTA performance on stained, high-contrast images but needs large labeled datasets, expensive hardware, and fails under uneven illumination. This study presents a low-cost, lightweight, annotation-free segmentation method by introducing one-time calibration-assisted unsupervised framework adaptable across imaging modalities and image type. The framework determines background via spatial standard deviation from the local mean. Uncertain pixels are resolved using fuzzy logic, cumulative squared shift of nodal intensity, statistical features, followed by post-segmentation denoising calibration which is saved as a profile for reuse until noise pattern or object type substantially change. The program runs as a script or graphical interface for non-programmers. The method was rigorously evaluated using \textit{IoU}, \textit{F1-score}, and other metrics, with statistical significance confirmed via Wilcoxon signed-rank tests. On unstained brightfield myoblast (C2C12) images, it outperformed \textit{Cellpose 3.0} and \textit{StarDist}, improving IoU by up to 48\% (average IoU = 0.43, F1 = 0.60). In phase-contrast microscopy, it achieved a mean IoU of 0.69 and an F1-score of 0.81 on the \textit{LIVECell} dataset ($n = 3178$), with substantial expert agreement ($κ> 0.75$) confirming cross-modality robustness. Successful segmentation of laser-affected polymer surfaces further confirmed cross-domain robustness. By introducing the \textit{Homogeneous Image Plane} concept, this work provides a new theoretical foundation for training-free, annotation-free segmentation. The framework operates efficiently on CPU, avoids cell staining, and is practical for live-cell imaging and biomedical applications.

Semi-Unsupervised Microscopy Segmentation with Fuzzy Logic and Spatial Statistics for Cross-Domain Analysis Using a GUI

TL;DR

This work addresses the challenge of segmenting unstained live cells in low-contrast bright-field microscopy without annotations or training data. It introduces the Homogeneous Image Plane concept and a one-time calibration–driven unsupervised pipeline that combines SSDLM-based homogeneity measures, fuzzy inference, Moran’s I, and an adjusted variogram to separate background texture from cell morphology. The approach yields cross-domain robustness across bright-field, phase-contrast, and laser-trace data, achieving up to a 48% IoU gain over state-of-the-art methods while running on CPU and offering a GUI for non-programmers. These results establish a practical, training-free alternative for live-cell imaging analysis with significant potential for regenerative medicine and clinical microscopy workflows, while also outlining avenues for automation and broader modality validation.

Abstract

Brightfield microscopy of unstained live cells is challenging due to low contrast, dynamic morphology, uneven illumination, and lack of labels. Deep learning achieved SOTA performance on stained, high-contrast images but needs large labeled datasets, expensive hardware, and fails under uneven illumination. This study presents a low-cost, lightweight, annotation-free segmentation method by introducing one-time calibration-assisted unsupervised framework adaptable across imaging modalities and image type. The framework determines background via spatial standard deviation from the local mean. Uncertain pixels are resolved using fuzzy logic, cumulative squared shift of nodal intensity, statistical features, followed by post-segmentation denoising calibration which is saved as a profile for reuse until noise pattern or object type substantially change. The program runs as a script or graphical interface for non-programmers. The method was rigorously evaluated using \textit{IoU}, \textit{F1-score}, and other metrics, with statistical significance confirmed via Wilcoxon signed-rank tests. On unstained brightfield myoblast (C2C12) images, it outperformed \textit{Cellpose 3.0} and \textit{StarDist}, improving IoU by up to 48\% (average IoU = 0.43, F1 = 0.60). In phase-contrast microscopy, it achieved a mean IoU of 0.69 and an F1-score of 0.81 on the \textit{LIVECell} dataset (), with substantial expert agreement () confirming cross-modality robustness. Successful segmentation of laser-affected polymer surfaces further confirmed cross-domain robustness. By introducing the \textit{Homogeneous Image Plane} concept, this work provides a new theoretical foundation for training-free, annotation-free segmentation. The framework operates efficiently on CPU, avoids cell staining, and is practical for live-cell imaging and biomedical applications.

Paper Structure

This paper contains 43 sections, 15 equations, 9 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unsupervised Learning for Image Segmentation Based on CNN
  4. Bayesian Statistical Model
  5. Soft Computing Based Approaches
  6. Unsupervised Domain Adaptation for Microscopy Images
  7. Hyperspectral Image Segmentation
  8. Classical and Hybrid Approaches
  9. Methodological Background & Technical Discussion
  10. Introducing Image Plane
  11. Notion of Homogeneous Image Plane & Homogeneity
  12. Homogeneous Image Plane
  13. Homogeneity of Region of Interest (ROI) in Image
  14. Other Spatial Statistical Metric Used
  15. Moran’s I
  16. ...and 28 more sections

Figures (9)

  • Figure 1: schematic diagram of end-to-end workflow
  • Figure 2: Membership functions $\mu_{dark}$, $\mu_{gray}$, and $\mu_{bright}$ for black, gray, and white regions.
  • Figure 3: Image after masking based on lower bound and upper bound where pink color denotes the uncertainty regions (left), Segmented image with Noise (Right)
  • Figure 4: Left: Hyperparameter-Tuning Window, Right: Post-Segmentation Denoising Window
  • Figure 5: Two different outputs generated by two hyperparameter tunings
  • ...and 4 more figures