Enhanced Feature Pyramid Network Using a Binary-Encoded Genetic Algorithm for Brain Tumor Segmentation

Abstract

Manually segmentation of brain tumors in magnetic resonance imaging (MRI) is time-consuming and requires years of professional experience and clinical knowledge. Automatic brain tumor segmentation is essential for diagnosis, tumor treatment planning, and patient monitoring using MRI data. For this task, the hyperparameters of deep learning models are often tuned manually, which makes the process less effective and more compute-intensive. This study proposed an Enhanced Feature Pyramid Network (FPN) optimized by Binary Encoding Genetic Algorithm (GA) for automatic brain tumor segmentation. The proposed framework combined a multi-scale FPN backbone and an evolutionary optimization method in which the hyperparameters of the architecture and training process are represented as binary chromosomes and evolved according to the segmentation performance. The genetic algorithm automatically tuned the parameters of the network like the depth, filter size, dropout rate, learning rate, optimizer type, activation function, and loss function by maximizing the average Dice coefficient. The experimental results on the BraTS 2020 dataset showed that the GA-optimized FPN outperforms the baseline FPN with Dice scores of 0.9703, 0.9594 and 0.9408 in the Whole Tumor, Tumor Core, and Enhancing Tumor regions, respectively, with the number of trainable parameters reduced by 93.7%. The findings demonstrated that binary-encoded evolutionary optimization can be used to boost segmentation accuracy while yielding models with substantially reduced trainable parameters and training time, showing improving parameter efficiency