A novel MRI-based deep learning imaging biomarker for comprehensive assessment of the lenticulostriate artery-neural complex

Yan Song, Yunlong Jin, Jianguo Wei, Jiajia Wang, Zhong Zheng, Ying Wang, Ru Zeng, Weiping Lu, Bingcang HuangBMC Medical Imaging 2025 · DOI

Abstract

The assessment of the lenticulostriate artery-neural complex (LNC) is critical for understanding subcortical vascular health, yet current multi-modal imaging protocols are often clinically impractical due to time and resource constraints. This study introduces a deep learning-based imaging biomarker capable of synthesizing structural, perfusion, and diffusion information from standard T1-weighted MRI (T1WI) scans. Utilizing a ResNet18 architecture, the model was trained to correlate T1WI features with physiological ground truths derived from Arterial Spin Labeling (ASL) and Diffusion Tensor Imaging (DTI). The framework achieved 100% accuracy in classifying both arterial morphology and patient disease status within a cohort of 79 individuals, suggesting that standard structural scans contain latent information sufficient for comprehensive LNC assessment.

Introduction

The lenticulostriate artery-neural complex (LNC) is a vital but fragile network in the deep brain. These tiny arteries, averaging just 0.36 mm in diameter, supply the basal ganglia—a region highly susceptible to strokes and hemorrhages. In clinical practice, evaluating the LNC is a fragmented process. Radiologists typically rely on Flow-Sensitive Black-Blood (FSBB) MRI to visualize artery structure, Arterial Spin Labeling (ASL) to measure cerebral blood flow (perfusion), and Diffusion Tensor Imaging (DTI) to assess nerve fiber integrity.

Performing this full battery of specialized scans is rarely feasible in routine clinical settings. This research addresses the need for a "one-stop" biomarker—a method to extract multi-modal physiological information from a single, standard structural MRI, thereby streamlining the diagnostic process for subcortical vascular diseases.

Methodology

The researchers proposed a deep learning framework based on the ResNet18 architecture to bridge the gap between structural anatomy and physiological function. The pipeline consists of several key stages:

  1. Automated Segmentation: Using advanced neuroimaging tools (FreeSurfer and ANTsPy), the team automatically identified three key regions of interest (ROIs) in the basal ganglia: the caudate nucleus, globus pallidus, and putamen.
  2. Correlation Training: The model was first trained to extract features from T1-weighted images that correlate with physiological "ground truth" data. Specifically, it targeted Cerebral Blood Flow (CBF) from ASL and Fractional Anisotropy (FA) from DTI. Root Mean Squared Error (RMSE) was utilized to measure the fidelity of these deep learning features against actual physiological values.
  3. Fine-Tuning and Classification: The pre-trained model was then tasked with two clinical classification challenges: identifying the number of LSA "roots" (arterial branches) and categorizing patients into three groups based on their clinical profile (healthy, diabetes only, or comorbid diabetes and hypertension).

Dataset and Study Population

The study utilized a retrospective dataset of 79 patients aged 50 and older who had undergone a complete multi-modal MRI battery. The selection process ensured high image quality and the presence of all necessary sequences for ground-truth comparison.

Figure 1: Flowchart of patient selection and clinical group categorization.

As shown in the flowchart, the final eligible sample of 79 patients was divided into three distinct clinical cohorts: 25 patients with both hypertension and diabetes, 27 with diabetes only, and 27 healthy controls without either condition.

Key Results

The deep learning model demonstrated a remarkable ability to extract physiological signatures from structural data. The RMSE loss relative to mean physiological values (FA and CBF) was exceptionally low, ranging from 0.154% to 0.213%. This indicates that the features extracted by the neural network are highly representative of actual blood flow and tissue integrity.

Figure 2: Representative Arterial Spin Labeling (ASL) map used as ground truth for training the perfusion-based features of the biomarker.

In the classification tasks, the model achieved 100% accuracy within this cohort for both identifying LSA root counts and distinguishing between the three disease categories. Visualization using t-SNE (t-distributed Stochastic Neighbor Embedding) further confirmed these results, showing clear, distinct clusters for different LSA counts and disease states, proving the robustness of the extracted features.

Discussion and Insights

The core finding of this study is that standard structural MRIs contain subtle "deep features" reflecting underlying vascular and neurological health that are invisible to the human eye. By treating the LNC as a unified, interactive entity, the model recognizes that damage to the LSA (vascular) is inextricably linked to changes in the surrounding gray matter (neural).

Analysis of the convolutional layers revealed a logical progression in feature extraction: shallow layers identified simple edges and textures, while deeper layers captured complex, abstract representations of brain health. This suggests the model is truly learning the physiological markers of disease rather than just memorizing image patterns.

Limitations and Future Directions

While the results are promising, the authors note several limitations. The small, single-center sample size means the 100% accuracy may not yet be generalizable to broader, more diverse populations. Additionally, the study lacked a "hypertension only" group, and the 2D measurement of ROIs could be improved with high-resolution 3D imaging. Future work will focus on validating the model with larger, multi-center datasets and exploring its application in cerebral small vessel disease and vascular cognitive impairment.

Summary updated

BibTeX
@article{song2025novel,
  author = {Yan Song and Yunlong Jin and Jianguo Wei and Jiajia Wang and Zhong Zheng and Ying Wang and Ru Zeng and Weiping Lu and Bingcang Huang},
  title = {{A novel MRI-based deep learning imaging biomarker for comprehensive assessment of the lenticulostriate artery-neural complex}},
  journal = {BMC Medical Imaging},
  year = {2025},
  doi = {10.1186/s12880-025-01676-3},
}
Figure