Improving the performance of EM and K-means algorithms for breast lesion segmentation

Abstract

Aims: Breast cancer is the most common type of cancer in women and accounts for a large portion of cancer-related deaths. As in the other types of cancer, the prevention and early diagnosis of breast cancer gain importance day after day. For this purpose, the artificial intelligence-based decision support systems become popular in recent years. In this study, an automatic breast lesion segmentation process is proposed to detect breast lesions in the images taken with magnetic resonance imaging (MRI) protocol. 
 Methods: Two most popular segmentation methods: expectation maximization (EM) and K-means algorithms are used to determine the region of breast lesions. Furthermore, superpixel based fuzzy C-means (SPFCM) algorithm is applied after EM and K-means methods to improve the lesion segmentation performance.
 Results: The proposed methods are evaluated on the private database constructed by the authors with ethical permission. The performances of the utilized methods are analyzed by comparing the lesion areas determined by a radiologist (ground-truth) and areas that are achieved by automatic segmentation algorithms. 
 Conclusion: Dice coefficient, Jaccard index (JI), and area under curve (AUC) metrics are calculated for performance comparison. According to the simulation results, EM, K-means, EM+SPFCM, and K-means+SPFCM methods provides good segmentation performance on breast MRI database. The best segmentation results are obtained by using EM+SPFCM hybrid method. The results of the EM+SPFCM method are 0,8711, 0,8979, and 0,9981 for JI, Dice, and AUC, respectively.