A new micro-genetic crossover operator for effective training of medical neural networks

Abstract

Clinical data classification became critical for medical decision-making support systems. While classification methods like multi-layer perceptrons have significantly contributed to disease diagnosis, their performance is often limited by the backpropagation algorithm’s susceptibility to local minima, slow convergence rates, and sensitivity to hyperparameter settings. Given that neural network training is an NP-Hard problem, optimizing weight and biases efficiently remains a significant challenge. Alternative strategies are needed to enhance classification performance while overcoming backpropagation’s drawbacks without imposing excessive computational costs. Metaheuristic approaches have gained attention for their adaptability and ability to find high-quality solutions with reasonable computational effort, even with NP-Hard problems. However, their application to neural networks often lacks a specific mechanism for updating weights and biases, limiting training efficiency. This study introduces the Crossed Knowledge Crossover operator, a new mechanism designed for the micro-genetic algorithm. Unlike traditional crossover methods, our operator combines entire blocks of weights and biases from distinct perceptron configurations, promoting the integration of features learned by different networks and increasing the likelihood of discovering superior configurations that enhance classification quality. Experiments include 13 well-known medical datasets, 31 state-of-the-art algorithms and 12 different versions of the Micro-Genetic Algorithm. The performance of each method is assessed by evaluating five classification metrics and comparing their time consumption. Experimental evaluations demonstrate that our approach outperforms established algorithms from the literature, revealing solutions that improve classification quality across at least 8 out of 13 datasets, with accuracy gains exceeding 5% regarding traditional methods. Highlighted results are obtained when classifying Breast Cancer and Parkinson’s disease data, reaching an accuracy of 98% and 97% respectively, overcoming most of the other approaches by a noticeable difference.