Search

Breast cancer is a significant health concern due to its aggressive nature and high mortality rates. Early detection is crucial to improving patient outcomes. Thermography, a non-invasive and cost-effective method, utilizes heat from the breast surface to detect abnormalities, with the Database for Mastology Research (DMR) leveraging infrared images for research purposes. Deep learning (DL), particularly convolutional neural networks (CNNs), shows promise in accurately classifying breast cancer images. However, CNNs face challenges with hyperparameters. To address these challenges, this article proposes a new DL model, the optimized DenseNet model (LFR-COA-DenseNet121-BC), incorporating a boosted metaheuristic algorithm called LFR-COA. This algorithm, a developed version of the Coati Optimization Algorithm (COA), integrates Random opposition-based learning (ROB), Brownian motion, and Lévy Flight (LF) schemes. The proposed model achieved impressive results, accurately classifying 99.97% of the test set. Comparison with established models such as VGG16, VGG19, DenseNet201, InceptionV3, Xception, and MobileNet revealed superior performance of the LFR-COA-DenseNet121-BC model, with 99.97% accuracy, 99.96% sensitivity, and 99.9% specificity. Further comparison with COA-DenseNet121 highlighted the superiority of LFR-COA-DenseNet-BC. In addition, LFR-COA was evaluated at the IEEE Congress on Evolutionary Computation held in 2022 (CEC 2022), and real-world medical scenarios showcased the effectiveness of LFR-COA compared to existing optimization algorithms. LFR-COA consistently outperformed the original COA algorithm and other well-known counterparts in various statistical, convergence, and diversity measures, affirming its efficacy in breast cancer classification. • CEC'22 test suite is utilized for verification of LFR-COA performance. • The LFR-COA algorithm optimizes various hyperparameters of the DenseNet121 model. • LFR-COA algorithm is proposed for solving global optimization problems. • LFR-COA-DenseNet121-BC: a new diagnostic model for breast cancer using DMR-IR dataset. • The proposed model surpasses other SOTA methods. [ABSTRACT FROM AUTHOR]

Feature selection (FS) represents an optimization problem that aims to simplify and improve the quality of highly dimensional datasets through selecting prominent features and eliminating redundant and irrelevant data to classify results better. The goals of FS comprise dimensionality reduction and enhancing the classification accuracy in general, accompanied by great significance in different fields like data mining applications, pattern classification, and data analysis. Using powerful optimization algorithms is crucial to obtaining the best subsets of information in FS. Different metaheuristics, such as the Sooty Tern Optimization Algorithm (STOA), help to optimize the FS problem. However, such kind of techniques tends to converge in sub-optimal solutions. To overcome this problem in the STOA, an improved version called mSTOA is introduced. It employs the balancing exploration/exploitation strategy, self-adaptive of the control parameters strategy, and population reduction strategy. The proposed approach is proposed for solving the FS problem, but also it has been validated over benchmark optimization problems from the CEC 2020. To assess the performance of the mSTOA, it has also been tested with different algorithms. The experiments in terms of FS provide qualitative and quantitative evidence of the capabilities of the mSTOA for extracting the optimal subset of features. Besides, statistical analyses and no-parametric tests were also conducted to validate the result obtained by the mSTOA in optimization. • An enhanced algorithm called the mSTOA that employs three strategies is proposed. • mSTOA efficiency and performance are verified on several benchmarks. • CEC'20 test suite are used for algorithm validation. • mSTOA is proposed as an alternate feature selection approach. [ABSTRACT FROM AUTHOR]