An improved weighted mean of vectors optimizer for multi-threshold image segmentation: case study of breast cancer

Abstract

Women are commonly diagnosed with breast cancer (BC), and early detection can significantly increase the cure rate. This study suggested a multi-threshold image segmentation (MTIS) technique for dividing BC histological slice images to assist in identifying lesions and boost diagnostic effectiveness. The selection of the threshold combination, a challenging combinatorial optimization problem, is the key to the MTIS approach. To enhance the MTIS method, a variant of INFO (BQINFO) is proposed to optimize the threshold combination selection procedure. BQINFO is constructed by introducing the barebones mechanism (BM) and quasi-opposition-based learning (QOBL) to INFO and addressing its slow convergence and weakness in local stagnation. To evaluate the optimization performance of BQINFO and the positive impact influence of introducing QOBL and BM to the original INFO for the acceleration of convergence speed and the solution of local stagnation, a series of comparative experiments were carried out using CEC2014 and CEC2021. The comprehensive results and comparisons obtained from the optimization indicators indicate the outstanding performance of BQINFO in overcoming the slow convergence and local stagnation problems when dealing with benchmark function problems. Besides, to further validate BQINFO's performance optimization of threshold combination selection, this paper performed an MTIS experiment with Rényi's entropy as the objective function on BSD500 images and BC histological slice images, respectively, providing qualitative and quantitative analysis with three evaluation metrics, FSIM, PSNR, and SSIM at low and high threshold levels. Ultimately, the experimental results demonstrate that BQINFO performs better and finds the optimal combination of thresholds faster than other comparison algorithms for both low and high threshold levels.