Transferring Ultrahigh-Field Representations for Intensity-Guided Brain Segmentation of Low-Field Magnetic Resonance Imaging

TL;DR

The paper tackles the bottleneck of limited access to ultrahigh-field 7T MRI for accurate brain segmentation by introducing a KD-driven framework that transfers 7T-like feature representations from a teacher trained on paired 7T data to a 7T-absent, low-field domain. A knowledge keeper network (KKN) learns to convert 3T features into 7T-like representations, while an adaptive fusion module (AFM) assimilates these cues into LF features to guide arbitrary segmentation models. Key contributions include a feature-level distillation pipeline with a 3D/2D training strategy, the AFM’s knowledge aggregation and channel-wise refinement, and extensive demonstrations of improved tissue- and whole-brain segmentation across IBSR and MALC datasets, plus ablation analyses. The work enables scalable, high-contrast segmentation in settings lacking 7T data, with potential extensions to tumor segmentation and super-resolution in clinical imaging workflows.

Abstract

Ultrahigh-field (UHF) magnetic resonance imaging (MRI), i.e., 7T MRI, provides superior anatomical details of internal brain structures owing to its enhanced signal-to-noise ratio and susceptibility-induced contrast. However, the widespread use of 7T MRI is limited by its high cost and lower accessibility compared to low-field (LF) MRI. This study proposes a deep-learning framework that systematically fuses the input LF magnetic resonance feature representations with the inferred 7T-like feature representations for brain image segmentation tasks in a 7T-absent environment. Specifically, our adaptive fusion module aggregates 7T-like features derived from the LF image by a pre-trained network and then refines them to be effectively assimilable UHF guidance into LF image features. Using intensity-guided features obtained from such aggregation and assimilation, segmentation models can recognize subtle structural representations that are usually difficult to recognize when relying only on LF features. Beyond such advantages, this strategy can seamlessly be utilized by modulating the contrast of LF features in alignment with UHF guidance, even when employing arbitrary segmentation models. Exhaustive experiments demonstrated that the proposed method significantly outperformed all baseline models on both brain tissue and whole-brain segmentation tasks; further, it exhibited remarkable adaptability and scalability by successfully integrating diverse segmentation models and tasks. These improvements were not only quantifiable but also visible in the superlative visual quality of segmentation masks.

Figures (13)

• Figure 1: Conceptual differences between conventional learning-based 3T-to-7T methods using (a) a 7T-like image as the guidance, (b) directly matched features as the guidance, and (c) the proposed method. Here, $\mathcal{E}$ and $\mathcal{G}$ denote an encoder and decoder, respectively.
• Figure 1: Visualization of the averaged feature maps extracted or transformed at the lowest level (Level 1) by the teacher encoder or the knowledge keeper ( i.e., our KKN).
• Figure 1: Comparison of brain tissue segmentation results of 3D U-Net between the image-level and feature-level UHF guidance on the IBSR dataset.
• Figure 1: Visualization of the lowest-level feature maps from our AFM trained on the IBSR dataset. In AFM, guided features are obtained by adding low-field features and UHF guidance. Each column refers to a different baseline model that applied our method.
• Figure 2: Schematic overview of our method. (a) A teacher network $\mathcal{T}$ is trained to extract meaningful UHF representations via 7T image reconstruction. (b) The guide blocks $\mathcal{B}$ within our KKN $\mathcal{K}$ learn how to map a 3T image $\mathbf{X}_p$ into UHF feature representations. (c) The AFMs $\mathcal{F}$ integrate UHF feature representations from the pre-trained KKN $\mathcal{K}$ with LF features $\mathbf{F}^\mathcal{S}_i$ extracted by a segmentation encoder $\mathcal{S}_\mathcal{E}$ and then transfer such fused features to the segmentation decoder $\mathcal{S_G}$ to produce 7T-guided segmentation masks $\mathbf{\widehat{S}}$.