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

Mohammad Hamghalam, Baiying Lei, Tianfu WangAAAI 2020 · DOI

Abstract

Accurate segmentation of brain tumors from MRI scans is often hindered by low tissue contrast and overlapping intensity distributions between tumor sub-regions and healthy tissue. This paper introduces a multi-stage Attention-GAN framework designed to synthesize High Tissue Contrast (HTC) images. By transforming standard MRI sequences into a domain with non-overlapping intensity distributions, the proposed method enhances the performance of downstream segmentation networks. Evaluated on the BraTS 2018 dataset, the approach demonstrates improved Dice scores and the potential to substitute missing MRI modalities in clinical workflows.

Introduction

In neuro-oncology, Magnetic Resonance Imaging (MRI) is essential for diagnosing and monitoring gliomas. However, standard sequences like FLAIR, T2, and T1c often suffer from "intensity overlap," where different tumor components—such as edema, non-enhancing tumor, and the enhancing core—share similar pixel values. This lack of contrast makes precise boundary mapping difficult for both radiologists and automated AI models, complicating surgical and radiation therapy planning.

Methodology: Multi-Stage Attention-GAN

The core innovation of this work is a framework that synthesizes high-contrast versions of MRI scans before performing segmentation. Instead of solely optimizing the segmentation model, the researchers focus on improving the input data quality.

High Tissue Contrast (HTC) Synthesis

The framework utilizes a CycleGAN-based architecture to translate standard MRI scans into a synthetic "High Tissue Contrast" (HTC) domain. In this target domain, the intensity distributions of the tumor regions and healthy tissues are forced to be distinct, effectively eliminating overlap.

Overview of the HTC synthesis framework and the regions of interest (WT, TC, ET).

Multi-Stage Pipeline and Attention Mechanism

To handle the hierarchical nature of tumor sub-regions, the process is divided into three stages:

  1. Stage I: Focuses on the Whole Tumor (WT) using FLAIR scans.
  2. Stage II: Focuses on the Tumor Core (TC) using T2 scans, operating on a cropped region defined by Stage I.
  3. Stage III: Focuses on the Enhancing Tumor (ET) using T1c scans, further refining the focus based on Stage II.

An unsupervised attention block is integrated into the GAN generator. This block produces a mask that directs the model to modify only the tumorous regions while preserving the healthy brain anatomy, thereby reducing artifacts and noise.

The three-stage pipeline for sequential segmentation of WT, TC, and ET using synthetic HTC images.

Experimental Setup and Results

The model was trained and validated using the BraTS 2018 dataset, which contains multi-modal MRI scans (FLAIR, T1, T1c, T2) with expert annotations.

Performance Gains

Integrating synthetic HTC images into standard segmentation networks (e.g., U-Net, FC-DenseNet) led to consistent improvements in Dice scores:

  • Whole Tumor (WT): +0.8%
  • Tumor Core (TC): +0.6%
  • Enhancing Tumor (ET): +0.5%

Modality Replacement and Efficiency

A significant finding was the ability to replace real T1 MRI sequences with synthetic versions without a loss in segmentation accuracy. This could potentially reduce patient time in the MRI scanner. Furthermore, the synthesis of a 2D HTC slice is highly efficient, taking approximately 27 milliseconds, making it viable for real-time clinical use.

Conclusion and Future Work

The Multi-Stage Attention-GAN offers a robust solution to the intensity overlap problem in brain tumor MRI. By prioritizing data clarity through synthesis, the framework simplifies the segmentation task. While the current 2D implementation shows promise, future work aims to extend the framework to 3D volumetric synthesis to improve spatial consistency. Additionally, addressing the sensitivity of the model to target distribution hyperparameters remains a key area for refinement.

Summary updated

BibTeX
@inproceedings{hamghalam2020high,
  author = {Mohammad Hamghalam and Baiying Lei and Tianfu Wang},
  title = {{High Tissue Contrast MRI Synthesis Using Multi-Stage Attention-GAN for Segmentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i04.5825},
}
Figure