Multi-level threshold segmentation framework for breast cancer images using enhanced differential evolution

Breast cancer has replaced lung cancer as the most prevalent malignancy threatening human health. Early breast screening can help improve treatment success and reduce the risk of death. The analysis and diagnosis of breast cancer real images by computer-aided technology is the key link to early diagnosis. High-quality medical segmentation images can improve the accuracy of lesion area detection. This study used a multi-level threshold image segmentation framework based on novel differential evolution, two-dimensional Kapur's entropy, and the two-dimensional histogram to improve the efficiency of subsequent image analysis and diagnosis. We proposed an enhanced differential evolution in the framework based on the roundup search, the elite lévy-mutation, and the decentralized foraging strategy to explore the optimal thresholds. In this study, the enhanced differential evolution was compared to state-of-the-art methods for benchmark function experiments and breast cancer image segmentation experiments. It is shown that the proposed threshold search method accelerates convergence and reduces the problem of premature convergence. Quantitative results demonstrate that the proposed method can achieve an average peak signal-to-noise ratio and feature similarity index of 21.231 and 0.951, respectively, at the 5-level threshold, which is better than other methods. As a result, the proposed multi-level threshold image segmentation model can provide quality samples for subsequent image analysis and classification.