Optimising Deep Feature Maps using Genetic Algorithm for Efficient Breast Mass Classification in Mammogram Images

Abstract

Breast cancer is the most prominent cause of death of women worldwide, highlighting the importance of precise and effective diagnostic tools. Deep convolutional neural networks have already achieved substantial success. in medical image analysis but run into redundancy and computational cost problems with high-dimensional feature flattening. This paper provides a hybrid deep learning and metaheuristic approach to optimising deep feature maps prior to flattening. From the CBIS-DDSM breast cancer dataset, feature maps are drawn from the ReLU5 layer of AlexNet, and global average pooling (GAP) compresses them into a single representative value without changing size or losing information. A Genetic Algorithm (GA) is employed for optimisation at the channel level to determine the most informative and relevant channels to classify. The subset of features optimised is then fed into a softmax classifier to classify finally.  This method avoids redundant calculations, accelerates convergence, and decreases training time. Experiments show that the designed model performs better than the conventional AlexNet framework in both accuracy in classification and efficiency in computations. Notably, at 40 epochs, GA_AlexNet achieved a 72.51% accuracy, surpassing AlexNet by 8.44%.