Data Augmentation with Conditional Generative Adversarial Networks (cGANs) for Deep Learning-based Classification of Brain Tumor Magnetic Resonance Images
Sukunimetön, Mahnoor (2025-06-02)
Data Augmentation with Conditional Generative Adversarial Networks (cGANs) for Deep Learning-based Classification of Brain Tumor Magnetic Resonance Images
(02.06.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2025061064831
https://urn.fi/URN:NBN:fi-fe2025061064831
Tiivistelmä
Background and aims: Generative adversarial networks (GAN) have been popularly used in generating augmented data in medical imaging. However, a classical GAN model is prone to model collapse, class imbalance, and instability. The purpose of this study was to validate a deep learning (DL) algorithm that generated brain tumor and non-tumor images from magnetic resonance images (MRI) and to compare its performance with that of true brain tumor and true brain non-tumor images from MRI.
Materials and methods: This single-center, retrospective study included MRI brain tumor and healthy control images from a public repository. Datasets were divided into training (63%), validation (6%), and test (31%) sets, with stratification by presence and absence of brain tumor. A conditional-generative adversarial network was trained to produce brain tumor and non-tumor images. The generated images were trained on a modified U-Net CNN multiple times with different numbers of generated images, and their classification accuracy was evaluated from a separate set of unseen dataset of 2000 images. The Mann-Whitney U-test was used to estimate the statistical significance between generated and true images. The generated images are systematically evaluated by multiple evaluation metrics, such as the Inception Score (IS), Frechet Inception Distance (FID), Structural Similarity Index Measure (SSIM), Mean Squared Error (MSE), Dice Similarity Coefficient (DSC), and Peak Signal-to-Noise Ratio (PSNR).
Results: A total of 2000 MRI images were generated, having an equal number of brain tumor and non-tumor images. The CNN trained with 1000 true and 1500 generated images worked the best, achieving 92% accuracy, 90% sensitivity, and 85% specificity for the diagnosis of brain tumors. The generated images exhibited an IS, SSIM, MSE, DSE, FID, and average PSNR of 2.09, 0.16, 7609, 0.89, 0.64, and 28.78 dB, respectively.
Conclusion: The classification performance of convolutional neural networks (CNNs) increased when its training set was augmented with generated MRI brain tumor and non-tumor images, suggesting that synthetic images can serve as effective alternatives to real images in deep learning-based classification models.
Significance: The results highlight the potential of generative MRI images as viable alternatives to real MRI scans for CNN-based brain tumor classification. It addresses data scarcity, enhances model robustness, preserves patient privacy, and reduces costs, making AI-driven diagnostics more scalable and efficient.
Materials and methods: This single-center, retrospective study included MRI brain tumor and healthy control images from a public repository. Datasets were divided into training (63%), validation (6%), and test (31%) sets, with stratification by presence and absence of brain tumor. A conditional-generative adversarial network was trained to produce brain tumor and non-tumor images. The generated images were trained on a modified U-Net CNN multiple times with different numbers of generated images, and their classification accuracy was evaluated from a separate set of unseen dataset of 2000 images. The Mann-Whitney U-test was used to estimate the statistical significance between generated and true images. The generated images are systematically evaluated by multiple evaluation metrics, such as the Inception Score (IS), Frechet Inception Distance (FID), Structural Similarity Index Measure (SSIM), Mean Squared Error (MSE), Dice Similarity Coefficient (DSC), and Peak Signal-to-Noise Ratio (PSNR).
Results: A total of 2000 MRI images were generated, having an equal number of brain tumor and non-tumor images. The CNN trained with 1000 true and 1500 generated images worked the best, achieving 92% accuracy, 90% sensitivity, and 85% specificity for the diagnosis of brain tumors. The generated images exhibited an IS, SSIM, MSE, DSE, FID, and average PSNR of 2.09, 0.16, 7609, 0.89, 0.64, and 28.78 dB, respectively.
Conclusion: The classification performance of convolutional neural networks (CNNs) increased when its training set was augmented with generated MRI brain tumor and non-tumor images, suggesting that synthetic images can serve as effective alternatives to real images in deep learning-based classification models.
Significance: The results highlight the potential of generative MRI images as viable alternatives to real MRI scans for CNN-based brain tumor classification. It addresses data scarcity, enhances model robustness, preserves patient privacy, and reduces costs, making AI-driven diagnostics more scalable and efficient.