11 January 2024 in Computerized Medical Imaging and Graphics
Accurate brain tumor segmentation is critical for diagnosis and treatment planning, whereby multi-modal magnetic resonance imaging (MRI) is typically used for analysis. However, obtaining all required sequences and expertly labeled data for training is challenging and can result in decreased quality of segmentation models developed through automated algorithms. In this work, we examine the possibility of employing a conditional generative adversarial network (GAN) approach for synthesizing multi-modal images to train deep learning-based neural networks aimed at high-grade glioma (HGG) segmentation. The proposed GAN is conditioned on auxiliary brain tissue and tumor segmentation masks, allowing us to attain better accuracy and control of tissue appearance during synthesis. To reduce the domain shift between synthetic and real MR images, we additionally adapt the low-frequency Fourier space components of synthetic data, reflecting the style of the image, to those of real data. We demonstrate the impact of Fourier domain adaptation (FDA) on the training of 3D segmentation networks and attain significant improvements in both the segmentation performance and prediction confidence. Similar outcomes are seen when such data is used as a training augmentation alongside the available real images. In fact, experiments on the BraTS2020 dataset reveal that models trained solely with synthetic data exhibit an improvement of up to 4% in Dice score when using FDA, while training with both real and FDA-processed synthetic data through augmentation results in an improvement of up to 5% in Dice compared to using real data alone. This study highlights the importance of considering image frequency in generative approaches for medical image synthesis and offers a promising approach to address data scarcity in medical imaging segmentation.
7 March 2024 in Computer Methods and Programs in Biomedicine
As large sets of annotated MRI data are needed for training and validating deep learning based medical image analysis algorithms, the lack of sufficient annotated data is a critical problem. A possible solution is the generation of artificial data by means of physics-based simulations. Existing brain simulation data is limited in terms of anatomical models, tissue classes, fixed tissue characteristics, MR sequences and overall realism. We propose a realistic simulation framework by incorporating patient-specific phantoms and Bloch equations-based analytical solutions for fast and accurate MRI simulations. A large number of labels are derived from open-source high-resolution T1w MRI data using a fully automated brain classification tool. The brain labels are taken as ground truth (GT) on which MR images are simulated using our framework. Moreover, we demonstrate that the T1w MR images generated from our framework along with GT annotations can be utilized directly to train a 3D brain segmentation network. To evaluate our model further on larger set of real multi-source MRI data without GT, we compared our model to existing brain segmentation tools, FSL-FAST and SynthSeg. Our framework generates 3D brain MRI for variable anatomy, sequence, contrast, SNR and resolution. The brain segmentation network for WM/GM/CSF trained only on T1w simulated data shows promising results on real MRI data from MRBrainS18 challenge dataset with a Dice scores of 0.818/0.832/0.828. On OASIS data, our model exhibits a close performance to FSL, both qualitatively and quantitatively with a Dice scores of 0.901/0.939/0.937. Our proposed simulation framework is the initial step towards achieving truly physics-based MRI image generation, providing flexibility to generate large sets of variable MRI data for desired anatomy, sequence, contrast, SNR, and resolution. Furthermore, the generated images can effectively train 3D brain segmentation networks, mitigating the reliance on real 3D annotated data.
1 November 2024 in Computers in Biology and Medicine
In the field of medical imaging, high-resolution (HR) magnetic resonance imaging (MRI) is essential for accurate disease diagnosis and analysis. However, HR imaging is prone to artifacts and is not universally available. Consequently, low-resolution (LR) MRI images are typically acquired. Deep learning (DL)-based super-resolution (SR) techniques can transform LR images into HR quality. However, these techniques require paired HR-LR data for training the SR networks. This research aims to investigate the potential of simulated brain MRI data to train DL-based SR networks. We simulated a large set of anatomically diverse, voxel-aligned, and artifact-free brain MRI data at different resolutions. We utilized this simulated data to train four distinct DL-based SR networks and augment their training. The trained networks were then evaluated using real data from various sources. With our trained networks, we produced 0.7mm SR images from standard 1mm resolution multi-source T1w brain MRI. Our experimental results demonstrate that the trained networks significantly enhance the sharpness of LR input MR images. For single-source images, the performance of networks trained solely on simulated data is slightly inferior to those trained solely on real data, with an average structural similarity index (SSIM) difference of 0.025. However, networks augmented with simulated data outperform those trained on single-source real data when evaluated across datasets from multiple sources. Paired HR-LR simulated brain MRI data is suitable for training and augmenting diverse brain MRI SR networks. Augmenting the training data with simulated data can enhance the generalizability of the SR networks across real datasets from multiple sources.