3D MRI Image Synthesizing using Commix GAN
P. L. Chithra, S.D Dhivya · 2024
This paper proposes a novel approach using Generative Artificial Intelligence (GAI) to generate superficial brain tumor MRI sequences effectively. The GAI framework is trained on a brain tumor dataset obtained from magnetic resonance imaging (MRI) to comprehend the patterns, data distribution, mapping of data, and structure of input images. By incorporating Deep Convolution Generative Adversarial Network (DCGAN), Pix2PixGAN, and Wasserstein GAN (WGAN), the random noise from the dataset distribution is combined with an actual input image to generate a synthetic image. Performance metrics evaluate the accuracy and loss of the generated images, considering both content and style loss. Despite the impressive accuracy of 96.8%, the Deep convolution generative network performed better without the Style neural network and the Wasserstein network gives better accuracy with the Style neural network. Performance measures are utilized to evaluate the accuracy and error of produced images while taking into account the data distribution of authentic and artificially produced images from the Deep convolution network, Pix2Pix network, and Wasserstein network, as well as the content and style outcomes along with sensitivity, specificity, precision, and F1 score. It is imperative to incorporate all four MRI sequences (T1, T2, T1w, and Flair) to conduct a thorough assessment of patients with brain tumor, ensuring a comprehensive evaluation. However, the absence of these modalities due to latency constraints and image distortions can hinder the performance and accuracy of deep learning and machine learning models developed for classification and other purposes. This limitation complicates patient care and exacerbates class imbalances.