[Paper Review] Semi-Automatic Algorithm for Breast MRI Lesion Segmentation Using Marker-Controlled Watershed Transformation

Sulaiman Vesal, Andres Diaz‐Pinto+4 more|arXiv (Cornell University)|Dec 14, 2017

AI in cancer detection10 references3 citations

TL;DR

This paper proposes a semi-automatic marker-controlled watershed transformation (MCWT) method for breast MRI lesion segmentation, using the brightest pixels in a radiologist-defined region of interest (ROI) as markers to improve accuracy. The approach achieves a mean Dice coefficient of 0.7808 ± 0.1729 and Jaccard index of 0.6704 ± 0.2167 across 106 lesions, demonstrating robust performance on both benign and malignant cases with minimal manual input.

ABSTRACT

Magnetic resonance imaging (MRI) is an effective imaging modality for identifying and localizing breast lesions in women. Accurate and precise lesion segmentation using a computer-aided-diagnosis (CAD) system, is a crucial step in evaluating tumor volume and in the quantification of tumor characteristics. However, this is a challenging task, since breast lesions have sophisticated shape, topological structure, and high variance in their intensity distribution across patients. In this paper, we propose a novel marker-controlled watershed transformation-based approach, which uses the brightest pixels in a region of interest (determined by experts) as markers to overcome this challenge, and accurately segment lesions in breast MRI. The proposed approach was evaluated on 106 lesions, which includes 64 malignant and 42 benign cases. Segmentation results were quantified by comparison with ground truth labels, using the Dice similarity coefficient (DSC) and Jaccard index (JI) metrics. The proposed method achieved an average Dice coefficient of 0.7808$\pm$0.1729 and Jaccard index of 0.6704$\pm$0.2167. These results illustrate that the proposed method shows promise for future work related to the segmentation and classification of benign and malignant breast lesions.

Motivation & Objective

• To address the challenge of accurate and robust segmentation of breast MRI lesions with complex shapes, variable intensity distributions, and topological complexity.
• To reduce over-segmentation in watershed-based segmentation by using a simple, intensity-based marker selection strategy.
• To develop a method that requires minimal manual input while maintaining high segmentation accuracy for use in computer-aided diagnosis (CAD) systems.
• To evaluate the performance of the proposed MCWT approach on a diverse set of 106 breast MRI lesions, including both benign and malignant cases.

Proposed method

• The method applies contrast-limited adaptive histogram equalization (CLAHE) to enhance image contrast in subtraction images.
• A radiologist manually defines a region of interest (ROI) around the lesion, which is then used to select the brightest n pixels as internal markers.
• The watershed transformation is applied on the gradient magnitude image using the selected bright pixels as internal markers and the ROI boundary as external markers.
• The optimal number of markers (n=45) is determined through cross-validation using Dice and Jaccard indices.
• A binary segmentation mask is generated from the watershed output, with regions labeled based on marker proximity.
• The method is evaluated using Dice similarity coefficient (DSC) and Jaccard index (JI) against expert manual segmentation.

Experimental results

Research questions

• RQ1 Can a simple, intensity-based marker selection strategy improve the robustness and accuracy of watershed-based lesion segmentation in breast MRI?
• RQ2 How does the number of markers affect segmentation performance in terms of DSC and JCI across diverse lesion types?
• RQ3 Can the proposed MCWT method achieve reliable segmentation on both benign and malignant lesions with varying shapes and intensity distributions?
• RQ4 To what extent does the method reduce over-segmentation compared to conventional watershed approaches?
• RQ5 Can the method serve as an effective preprocessing step for subsequent lesion classification in CAD systems?

Key findings

• The proposed MCWT method achieved a mean Dice similarity coefficient (DSC) of 0.7808 ± 0.1729 across 106 lesions, indicating strong overlap with ground truth segmentation.
• The mean Jaccard index (JI) was 0.6704 ± 0.2167, reflecting good agreement between automated and manual segmentation.
• The optimal number of markers was found to be 45, balancing segmentation accuracy and over-segmentation control.
• The method performed well on medium to large lesions, with accurate boundary detection and minimal marginal errors.
• For cases with multiple disjoint lesions, the method struggled due to intensity overlap between lesions and healthy tissue, especially when the ROI was large.
• The approach demonstrated robustness and simplicity compared to conventional marker-based methods that rely on complex features like template matching or Gaussian mixture models.