Abstract
Accurate segmentation of Medial Temporal Lobe (MTL) subregions is critical for early Alzheimer’s disease (AD) diagnosis. However, standard T2-weighted MRI scans are often anisotropic, leading to "staircase" artifacts that hinder precise morphometric measurements. This paper presents a novel pipeline using Implicit Neural Representations (INR) to generate high-resolution, isotropic training data from low-resolution anisotropic atlases. By training a multi-modality nnU-Net on this super-resolved data, we achieve superior segmentation performance. Our model demonstrates an enhanced ability to distinguish between healthy aging and mild cognitive impairment (MCI) and provides more stable longitudinal measurements compared to traditional methods.
Introduction
The Medial Temporal Lobe (MTL) is among the first brain regions to exhibit pathology in Alzheimer’s disease. Measuring the volume and thickness of its subregions, such as the hippocampus and entorhinal cortex, serves as a vital biomarker for disease progression. While T2-weighted MRI offers high contrast for these structures, clinical scans are typically anisotropic—high resolution in-plane but thick through-plane. This anisotropy creates significant challenges for automated segmentation tools, which often struggle with the resulting "staircase" artifacts and low through-plane resolution.
Methodology
The proposed framework addresses these limitations through a two-stage process: isotropic atlas creation and segmentation model training.
Isotropic Atlas Creation via INR
Rather than relying on standard interpolation techniques that may introduce blurring, the authors utilize Implicit Neural Representation (INR). In this approach, an image is modeled as a continuous function rather than a discrete voxel grid. A neural network is trained to map 3D coordinates $(x, y, z)$ to voxel intensities and anatomical labels. This allows for querying the network at any arbitrary resolution, effectively upsampling the anisotropic atlas into a smooth, $0.4 \text{ mm}^3$ isotropic space.
Segmentation Model Training
The resulting high-resolution isotropic atlas is used to train a multi-modality nnU-Net. This model integrates both T1-weighted structural scans and T2-weighted high-contrast scans. By training on "perfectly" smooth isotropic labels, the model learns the underlying biological geometry of the MTL, allowing it to perform robustly even on standard anisotropic test images.
Experimental Setup and Datasets
The model was evaluated across several key datasets:
- Training Set: The ASHS atlas, consisting of 29 participants with expert manual labels.
- Status Discrimination Set: 200 participants from the Penn ADRC used to evaluate the model's ability to distinguish healthy controls (CU) from those with Mild Cognitive Impairment (MCI).
- Test-Retest Set: 30 pairs of scans from 26 participants to measure measurement stability over time.
Results
- Superior Discrimination: The isotropic model outperformed standard baselines in clinical tasks. For instance, the volume of the Left CA1 subfield achieved an Area Under the Curve (AUC) of 0.861, surpassing the anisotropic baseline.
- Enhanced Stability: In test-retest analysis, the isotropic model produced significantly more consistent measurements, showing lower "noise" (standard deviation of change) across nearly all subregions compared to traditional ASHS methods.
- Visual Fidelity: Qualitatively, the model preserved the continuity of small structures like the CA2 and CA3 subfields, which are frequently fragmented in models trained directly on anisotropic data.
Discussion and Conclusion
The core innovation of this work lies in prioritizing the geometric quality of training data over simple voxel counts. By using INR as a bridge to create high-quality training sets, the researchers developed a model that is both clinically effective and computationally efficient for inference. While current inference uses linear upsampling for speed, future iterations may incorporate generative models for even higher fidelity. This approach represents a significant step forward in the automated, reliable measurement of brain biomarkers for Alzheimer’s research.