Abstract:Magnetic resonance imaging (MRI) provides varying tissue contrast images of internal organs based on a strong magnetic field. Despite the non-invasive advantage of MRI in frequent imaging, the low contrast MR images in the target area make tissue segmentation a challenging problem. This paper demonstrates the potential benefits of image-to-image translation techniques to generate synthetic high tissue contrast (HTC) images. Notably, we adopt a new cycle generative adversarial network (CycleGAN) with an attention mechanism to increase the contrast within underlying tissues. The attention block, as well as training on HTC images, guides our model to converge on certain tissues. To increase the resolution of HTC images, we employ multi-stage architecture to focus on one particular tissue as a foreground and filter out the irrelevant background in each stage. This multi-stage structure also alleviates the common artifacts of the synthetic images by decreasing the gap between source and target domains. We show the application of our method for synthesizing HTC images on brain MR scans, including glioma tumor. We also employ HTC MR images in both the end-to-end and two-stage segmentation structure to confirm the effectiveness of these images. The experiments over three competitive segmentation baselines on BraTS 2018 dataset indicate that incorporating the synthetic HTC images in the multi-modal segmentation framework improves the average Dice scores 0.8%, 0.6%, and 0.5% on the whole tumor, tumor core, and enhancing tumor, respectively, while eliminating one real MRI sequence from the segmentation procedure.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the difficult tumor segmentation in brain magnetic resonance imaging (MRI) due to low tissue contrast. Specifically, the author proposes a new multi - stage Attention - GAN (Multi - Stage Attention - GAN) to generate high - tissue - contrast (HTC) MRI images, so as to improve the segmentation accuracy of glioma. ### Main problems 1. **Low - contrast problem**: The contrast of the target area in traditional MRI images is low, resulting in difficulties in tissue segmentation. 2. **Multi - modal data utilization**: How to effectively utilize multiple MRI sequences (such as FLAIR, T2, T1c) to improve segmentation performance. 3. **Quality of synthetic images**: The generated synthetic HTC images need to be as close as possible to the real images to improve the accuracy of segmentation. ### Solutions 1. **Multi - stage attention mechanism**: By introducing the attention mechanism, the model can focus on specific tissue areas, thereby generating high - contrast synthetic images. 2. **Cycle - consistency loss**: Using the CycleGAN framework, through forward and backward cycle - consistency losses, ensure the mapping relationship between the generated synthetic images and the real images. 3. **Staged processing**: Divide the segmentation task into multiple stages, and each stage focuses on a specific tissue area, gradually improving the segmentation accuracy. ### Specific methods 1. **Image generation module**: - Use CycleGAN to generate the mapping from MRI to HTC, and focus on specific tissue areas through the attention mechanism. - The generated HTC images are mapped back to MRI images in reverse to verify the quality of the generated images. 2. **Segmentation module**: - Use the generated HTC images for binary - classification segmentation to generate foreground and background segmentation maps. - Through the multi - stage structure, gradually process the segmentation tasks of different tissue areas. ### Experimental results 1. **Quality of synthetic HTC images**: - Evaluate the similarity between the generated HTC images and the real HTC images through K - S test and PSNR, SSIM indicators. - The results show that the generated HTC images perform well on multiple indicators, especially under the combination of multi - stage structure and attention mechanism. 2. **Segmentation performance**: - Multi - modal segmentation experiments were carried out on the BraTS 2018 data set, and the results show that using the generated HTC images can significantly improve the Dice coefficient of segmentation. - Specifically, for the whole tumor (WT), tumor core (TC) and enhanced tumor (ET), the segmentation performance using HTC images is improved by 0.8%, 0.6% and 0.5% respectively. ### Contributions 1. **Propose a new multi - stage Attention - GAN**, which can generate high - contrast MRI images and improve the accuracy of tumor segmentation. 2. **Reduce the gap between the source domain and the target domain through the multi - stage structure**, and improve the quality of synthetic images. 3. **The experiments on the BraTS data set verify the effectiveness of the method**, showing superior performance in multi - modal segmentation tasks. In conclusion, this paper effectively solves the segmentation problem caused by low contrast in MRI images by introducing the multi - stage attention mechanism and CycleGAN framework, providing a new technical means for the accurate diagnosis of glioma.

High Tissue Contrast MRI Synthesis Using Multi-Stage Attention-GAN for Glioma Segmentation

Brain Tumor Synthetic Segmentation in 3D Multimodal MRI Scans

A multi-task generative model for simultaneous post-contrast MR image synthesis and brainstem glioma segmentation

Improvement of Multiparametric MR Image Segmentation by Augmenting the Data With Generative Adversarial Networks for Glioma Patients

Synthesizing MR Image Contrast Enhancement Using 3D High-Resolution ConvNets

Medical image synthesis via conditional GANs: Application to segmenting brain tumours

Image synthesis in contrast MRI based on super resolution reconstruction with multi-refinement cycle-consistent generative adversarial networks

MRI Contrast Enhancement Synthesis Using Cascade Networks with Local Supervision

Contrast-enhanced MRI Synthesis from Non-Contrast MRI Using Attention CycleGAN

T1-contrast Enhanced MRI Generation from Multi-parametric MRI for Glioma Patients with Latent Tumor Conditioning

Magnetic Resonance Imaging Contrast Enhancement Synthesis Using Cascade Networks with Local Supervision

Nimg-65. Synthetic Mri Generation To Facilitate Autosegmentation Of Pediatric Brain Tumors In Limited Data Scenarios

Multi-Modality Generative Adversarial Networks with Tumor Consistency Loss for Brain MR Image Synthesis

MCMT-GAN: Multi-Task Coherent Modality Transferable GAN for 3D Brain Image Synthesis.

Attention-Based Multi-Scale Generative Adversarial Network for synthesizing contrast-enhanced MRI

Unpaired Multi-contrast MR Image Synthesis Using Generative Adversarial Networks.

Deep learning-based 3D MRI contrast-enhanced synthesis from a 2D noncontrast T2Flair sequence

Multi-Contrast MRI Segmentation Trained on Synthetic Images

Multi-modal brain tumor segmentation via conditional synthesis with Fourier domain adaptation

Attention-Guided Generative Adversarial Network to Address Atypical Anatomy in Modality Transfer

High-resolution MR Imaging Using Self-Supervised Parallel Network