Comparative Analysis of Unsupervised Algorithms for Breast MRI Lesion Segmentation

Sulaiman Vesal, Nishant Ravikumar, Stephan Ellman, Andreas MaierarXiv 2018 · arXiv

TL;DR

This paper evaluates three unsupervised clustering techniques—K-means, Gaussian Mixture Models (GMM), and Marker-Controlled Watershed Transformation (MCWT)—for segmenting breast lesions in MRI scans. The study finds that the MCWT approach consistently outperforms the others, offering a robust way to automate tumor boundary detection without the need for labor-intensive, manually labeled training data.

What Problem Does This Paper Address?

Accurate segmentation—the process of delineating the exact boundaries of a tumor—is a cornerstone of breast cancer diagnosis and treatment planning. While Magnetic Resonance Imaging (MRI) provides high-quality data, breast lesions are notoriously difficult to segment automatically. They often exhibit irregular shapes, complex topological structures, and intensity distributions that overlap significantly with healthy tissue.

Currently, many clinical assessments rely on manual measurements or supervised machine learning models. However, manual segmentation is slow, prone to human error, and highly subjective. Supervised models, while powerful, require massive amounts of expert-labeled "ground truth" data, which is expensive and time-consuming to produce. The authors tackle this by investigating unsupervised algorithms, which can identify patterns and boundaries in the data itself without prior training on labeled examples.

Methods

The researchers compared three distinct unsupervised strategies applied to 2D regions of interest (ROIs) within breast MRI slices:

  1. K-Means Clustering (KM): A classic algorithm that partitions pixels into $k$ clusters by assigning each pixel to the cluster with the nearest mean (centroid).
  2. Gaussian Mixture Model (GMM): A more sophisticated probabilistic approach that assumes the image data is composed of a mixture of several Gaussian distributions. It uses the Expectation-Maximization (EM) algorithm to iteratively refine cluster parameters.
  3. Marker-Controlled Watershed Transformation (MCWT): This method treats the image like a topographic map where pixel intensity represents "elevation." To prevent the common problem of "over-segmentation" (where the algorithm finds too many tiny regions), the authors used specific high-intensity pixels as "markers" to guide the flooding process that defines boundaries.

Before applying these algorithms, the team used Contrast-Limited Adaptive Histogram Equalization (CLAHE) to enhance image contrast, making the lesions stand out more clearly from the background.

Datasets

The study utilized a dataset of 106 lesions (59 malignant and 47 benign) identified from 80 female patients. The images were acquired using 1.5T and 3.0T Siemens scanners.

  • Ground Truth: Two expert radiologists with over 7 years of experience provided manual segmentations to serve as the gold standard for evaluation.
  • Confirmation: In all cases, the cancer status (benign vs. malignant) was confirmed via histopathology.

Key Insights and Ideas

  • The Power of Markers: The standard Watershed algorithm often fails due to noise, leading to a fragmented "broken glass" effect. By selecting the top $n$ highest-intensity pixels (finding 45 to be optimal) as markers, the authors successfully constrained the algorithm to focus on the actual tumor mass.
  • Handling Disjoint Areas: One of the biggest challenges in lesion segmentation is when a tumor appears as several disconnected spots in a 2D slice. The MCWT proved particularly adept at "connecting the dots" compared to K-means or GMM.
  • Clinical Viability: Because these methods are unsupervised, they are uniquely suited for real-time applications in clinical settings where a radiologist needs immediate feedback without waiting for a model to be trained on a specific local dataset.

Key Results

The quantitative evaluation used several metrics, including the Dice Similarity Coefficient (DSC) and Jaccard Index (JI). As shown in the results table, MCWT emerged as the clear winner across almost every category.

Table 1: Quantitative comparison of the three algorithms. MCWT shows the highest accuracy across DSC, JI, and Precision metrics.
  • MCWT achieved a DSC of 0.786, significantly higher than GMM (0.746) and K-Means (0.732).
  • Qualitative Success: Visually, MCWT produced masks that most closely resembled the radiologist's ground truth, especially in complex, irregular lesions.
Fig. 1. Comparison of segmentation results for different methods, column one is the lesions, second is the ground truth masks and column third, fourth and fifth are the segmentation output for GMM, KM, and MCWT.

Limitations and Future Work

The primary limitation acknowledged by the authors is that this study was conducted on 2D slices rather than full 3D MRI volumes. While 2D analysis is a vital first step, tumors are 3D structures, and segmenting them across multiple slices introduces further complexity.

Additionally, while MCWT performed best, it remains sensitive to noise and "soft edges" where the gradient between the tumor and healthy tissue is very subtle. The authors suggest that future work will focus on extending the MCWT algorithm to 3D segmentation and integrating it into a complete Computer-Aided Diagnosis (CAD) system that can both detect and segment lesions with minimal human intervention.

Summary updated

Abstract

Accurate segmentation of breast lesions is a crucial step in evaluating the characteristics of tumors. However, this is a challenging task, since breast lesions have sophisticated shape, topological structure, and variation in the intensity distribution. In this paper, we evaluated the performance of three unsupervised algorithms for the task of breast Magnetic Resonance (MRI) lesion segmentation, namely, Gaussian Mixture Model clustering, K-means clustering and a marker-controlled Watershed transformation based method. All methods were applied on breast MRI slices following selection of regions of interest (ROIs) by an expert radiologist and evaluated on 106 subjects' images, which include 59 malignant and 47 benign lesions. Segmentation accuracy was evaluated by comparing our results with ground truth masks, using the Dice similarity coefficient (DSC), Jaccard index (JI), Hausdorff distance and precision-recall metrics. The results indicate that the marker-controlled Watershed transformation outperformed all other algorithms investigated.
BibTeX
@misc{vesal2018comparative,
  author = {Sulaiman Vesal and Nishant Ravikumar and Stephan Ellman and Andreas Maier},
  title = {{Comparative Analysis of Unsupervised Algorithms for Breast MRI Lesion Segmentation}},
  year = {2018},
  eprint = {1802.08655},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.08655},
}

Loading related content...

Figure