Deep learning-based automated segmentation of brain tumors using synthetic MR images generated with DCGAN
Computer Standards and Interfaces, cilt.96, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 96
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.csi.2025.104054
- Dergi Adı: Computer Standards and Interfaces
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Linguistic Bibliography
- Anahtar Kelimeler: Brain tumors, Deep learning, GAN, K-means, Segmentation
- Süleyman Demirel Üniversitesi Adresli: Evet
Özet
Early detection of a brain tumor significantly increases the likelihood that treatment will begin in a timely manner. Because it is difficult to detect tumor tissue with visual inspection, the magnetic resonance (MR) imaging method was developed. The analysis of MR images largely dependent on the radiologist's experience and visual interpretation. The primary reason for this is that brain tumors can vary in form and size. Deep learning (DL)-based techniques have accelerated medical image segmentation research thanks to their self-learning capabilities. When large amounts of training data are presented, these methods can achieve high success rates. ImageNet, CIFAR10/100, PASCAL VOC, MS COCO, and BRaTS benchmark datasets are extensively used for brain tumor segmentation. However, the limited amount of data in these datasets restricts the performance of DL models. The outstanding performance of Generative Adversarial Networks (GAN) in the field of medical image generation has attracted the interest of academics in recent years. In the study, we present a deep learning model that creates synthetic brain MR images using a Deep Convolutional GAN (DCGAN). The BRaTS2018 dataset's FLAIR sequence training data has been utilized as input. After a certain number of epochs, the learning model generated realistic and high-quality brain MR images. The FID score was used to evaluate the performance of the GAN model. Tumor regions on the generated MR images have been segmented automatically using the K-means algorithm and produced a high-dimensional dataset of 782 images. The study examined to what extent synthetic MR images enhanced the tumor region segmentation performance of the UNet, ResUNet, ResNet50, VGG16, and VGG19 models. According to the findings of the study, the ResNet50 model outperformed the other DL models. In terms of model performance, accuracy improved from 98.99% to 99.26%, the dice coefficient score moved from 57.33% to 81.32%, and the IoU increased from 40.89% to 66.86%.