Unsupervised Detection of Fibrosis in Microscopy Images Using Fractals and Fuzzy c-Means Clustering

Sotiris K. Tasoulis, Ilias Maglogiannis, Vassilis P. PlagianakosIFIP 2012 · DOI · OpenAlex

Introduction

Idiopathic Pulmonary Fibrosis (IPF) is a chronic, progressive, and often lethal lung disorder with a median survival rate of approximately three years. Early and accurate diagnosis is critical for patient management. Currently, pathologists assess lung tissue by manually scoring microscopic images, a process that is inherently subjective, time-consuming, and prone to inter-observer variability.

While supervised machine learning methods (such as Neural Networks or SVMs) have been applied to this problem, they require large volumes of expertly labeled training data. This paper presents an unsupervised computer-assisted tool that quantifies disease severity based on the intrinsic mathematical properties of the images themselves, eliminating the need for prior human labeling.

Methodology

The researchers developed a multi-step pipeline to transform raw microscopy images into quantitative maps of disease progression.

Flowchart of the proposed unsupervised pipeline for fibrosis detection.
  1. Binarization: Grayscale microscopic images are converted into binary (black and white) images using Otsu’s method, effectively separating foreground tissue from the background.
  2. Window Segmentation: The binary image is partitioned into small $n \times n$ windows (tested at 5, 10, and 20 pixels).
  3. Fractal Dimension (FD) Extraction: Using the box-counting method, the system calculates the fractal dimension of each window. In biological contexts, FD serves as a mathematical index of "roughness" or structural complexity, which differs significantly between healthy and fibrotic tissue.
  4. Feature Vector Construction: Each window is represented by a vector containing its Fractal Dimension and mean grayscale intensity.
  5. Fuzzy c-Means (FCM) Clustering: FCM assigns a "membership level" to each window rather than a rigid classification. This "fuzzy" logic accounts for the biological reality that tissue often exists on a spectrum between healthy and diseased states.

Experimental Results

The study utilized a biological model of 6-8 week-old mice with Bleomycin-induced pulmonary fibrosis. Lung sections were stained with Hematoxylin and Eosin and imaged at 4x magnification at various intervals (7, 15, and 23 days post-injection).

Visual progression of image processing from raw microscopic input to final clustered output.

The system was tested using a 4-cluster configuration, which successfully mapped the tissue into four distinct categories: Severe Pathology (Red), Mild Pathology (Orange), Normal Lung (Green), and Background (Blue).

Key findings include:

  • Validation: The automated cluster percentages closely aligned with manual scores and annotations provided by expert pathologists.
  • Progression Tracking: The tool successfully quantified the increase in "Severe Pathology" over time. For instance, using a 5-pixel window size, severe pathology was measured at 10.89% at day 7, increasing to 15.01% by day 23.
  • Resolution Sensitivity: While all window sizes (5, 10, and 20 pixels) were effective, the choice of resolution slightly shifted the detected pathology percentages, highlighting the importance of scale in automated scoring.

Conclusion and Future Work

This research demonstrates that combining fractal dimension analysis with fuzzy clustering provides a robust, unsupervised method for quantifying lung fibrosis. By capturing the geometric complexity of tissue architecture, the system offers an objective alternative to manual scoring.

Future work will focus on refining the accuracy of the assessment by incorporating additional image properties and implementing post-processing techniques to filter out isolated "noisy" regions in the clustered output.

Summary updated

BibTeX
@inproceedings{tasoulis2012unsupervised,
  author = {Sotiris K. Tasoulis and Ilias Maglogiannis and Vassilis P. Plagianakos},
  title = {{Unsupervised Detection of Fibrosis in Microscopy Images Using Fractals and Fuzzy c-Means Clustering}},
  booktitle = {IFIP},
  year = {2012},
  doi = {10.1007/978-3-642-33409-2\_40},
}
Figure