MRI-based machine learning model: A potential modality for predicting cognitive dysfunction in patients with type 2 diabetes mellitus

Xu Zhi-gao, Lili Zhao, Lei Yin, Yan Liu, Ying Ren, Guoqiang Yang, Jinlong Wu, Feng Long Gu, Xuesong Sun, Hui Yang, Taisong Peng, Jinfeng Hu, Xiaogeng Wang, Minghao Pang, Qiong Dai, Guojiang ZhangFrontiers in Bioengineering and Biotechnology 2022 · DOI

This research paper, published in Frontiers in Bioengineering and Biotechnology, explores the intersection of medical imaging and artificial intelligence to address a growing complication of metabolic disease: cognitive decline.

TL;DR

Researchers developed a machine learning pipeline that uses MRI radiomics to predict cognitive dysfunction in Type 2 Diabetes (T2DM) patients. By extracting subtle texture features from standard FLAIR brain scans, their Logistic Regression model successfully distinguished between normal cognition, mild impairment, and dementia with high accuracy (AUCs up to 0.90).

What Problem Does This Paper Address?

Type 2 Diabetes Mellitus (T2DM) is a well-established risk factor for cognitive impairment, ranging from mild deficits to full-blown dementia. Early detection is critical for intervention, yet current clinical practice relies heavily on neuropsychological screening tests like the Montreal Cognitive Assessment (MoCA).

While useful, these tests are subjective, can have poor repeatability, and are influenced by a patient’s education level. On the other hand, standard MRI scans are often used to look for structural changes like atrophy or "mini-strokes" (lacunar infarcts), but these visible markers often appear only after significant damage has occurred. The authors sought to determine if "radiomics"—the extraction of thousands of quantitative features from medical images—combined with machine learning could identify subtle "brain texture" changes that precede gross structural loss.

Methods

The researchers employed a radiomics-based machine learning workflow:

  1. Imaging: They utilized 3T MRI scans, specifically focusing on Fluid Attenuated Inversion Recovery (FLAIR) sequences, from 122 T2DM patients.
  2. Segmentation: Radiologists manually delineated the cerebrum as the Region of Interest (ROI) layer-by-layer.
  3. Feature Extraction: Using the Radcloud platform, they extracted 1,409 radiomic features, including first-order statistics (intensity distributions), shape-based features, and textural features (spatial relationships between pixels).
  4. Feature Selection: To prevent overfitting (the "curse of dimensionality"), they used a three-step reduction process:
  5. Variance Thresholding: Removing features with low variance.
  6. SelectKBest: Identifying features with the strongest individual correlation to the target.
  7. LASSO (Least Absolute Shrinkage and Selection Operator): A regression analysis method that performs both variable selection and regularization to enhance prediction accuracy.
  8. Classification: They compared three common machine learning algorithms: k-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Logistic Regression (LR).

Datasets

The study used a primary dataset of 122 patients with T2DM, categorized based on their MoCA-B (Chinese version) scores:

  • Dementia (DM) group: n = 40
  • Mild Cognitive Impairment (MCI) group: n = 52
  • Normal cognitive state (N) group: n = 30

The data was split into a training set (n=98) and a validation set (n=24) using a 8:2 ratio.

Key Insights and Ideas

  • Texture Over Volume: The study suggests that the "texture" of the brain—the subtle arrangement and intensity variations of voxels—contains predictive signatures for cognitive decline that are invisible to the naked eye and independent of simple volume loss.
  • The Power of Simplicity: Interestingly, the Logistic Regression (LR) model outperformed more complex algorithms like SVM and KNN. This suggests that the relationship between the selected radiomic features and cognitive state may be largely linear, or that LR is more robust when dealing with smaller medical datasets.
  • Clinical Feasibility: By using only FLAIR sequences—a standard part of most clinical brain MRI protocols—this approach is more easily translatable to real-world hospital settings than methods requiring specialized sequences like Diffusion Tensor Imaging (DTI) or functional MRI (fMRI).

Key Results

The feature selection process successfully whittled down 1,409 variables to just 13 optimal radiomic discriminators.

In the validation cohort, the Logistic Regression (LR) model demonstrated superior performance across all categories:

  • Normal Cognition (N): AUC of 0.904
  • Mild Cognitive Impairment (MCI): AUC of 0.883
  • Dementia (DM): AUC of 0.831

The model showed high specificity (0.950 for the Normal group), meaning it was particularly good at correctly identifying patients without cognitive impairment.

Limitations and Future Work

The authors acknowledge several hurdles before this can be used in a clinic:

  • Sample Size: The study was limited to 122 patients from a single center. Larger, multi-center studies are required to ensure the model generalizes across different populations and MRI hardware.
  • Manual Segmentation: Delineating the cerebrum manually is time-consuming and subject to human error. Developing automated, AI-driven segmentation tools is a necessary next step.
  • Sequence Limitations: The study only used FLAIR. The authors suggest that combining radiomics from multiple sequences (T1, T2, and DTI) might provide a more holistic "fingerprint" of the diabetic brain.
  • Retrospective Design: As a retrospective study, it cannot definitively prove that these MRI changes predict future decline; prospective longitudinal studies are needed to confirm the model's prognostic value.

Summary updated

Abstract

<b>Background:</b> Type 2 diabetes mellitus (T2DM) is a crucial risk factor for cognitive impairment. Accurate assessment of patients' cognitive function and early intervention is helpful to improve patient's quality of life. At present, neuropsychiatric screening tests is often used to perform this task in clinical practice. However, it may have poor repeatability. Moreover, several studies revealed that machine learning (ML) models can effectively assess cognitive impairment in Alzheimer's disease (AD) patients. We investigated whether we could develop an MRI-based ML model to evaluate the cognitive state of patients with T2DM. <b>Objective:</b> To propose MRI-based ML models and assess their performance to predict cognitive dysfunction in patients with type 2 diabetes mellitus (T2DM). <b>Methods:</b> Fluid Attenuated Inversion Recovery (FLAIR) of magnetic resonance images (MRI) were derived from 122 patients with T2DM. Cognitive function was assessed using the Chinese version of the Montréal Cognitive Assessment Scale-B (MoCA-B). Patients with T2DM were separated into the Dementia (DM) group (<i>n</i> = 40), MCI group (<i>n</i> = 52), and normal cognitive state (N) group (<i>n</i> = 30), according to the MoCA scores. Radiomics features were extracted from MR images with the Radcloud platform. The variance threshold, SelectKBest, and least absolute shrinkage and selection operator (LASSO) were used for the feature selection. Based on the selected features, the ML models were constructed with three classifiers, k-NearestNeighbor (KNN), Support Vector Machine (SVM), and Logistic Regression (LR), and the validation method was used to improve the effectiveness of the model. The area under the receiver operating characteristic curve (ROC) determined the appearance of the classification. The optimal classifier was determined by the principle of maximizing the Youden index. <b>Results:</b> 1,409 features were extracted and reduced to 13 features as the optimal discriminators to build the radiomics model. In the validation set, ROC curves revealed that the LR classifier had the best predictive performance, with an area under the curve (AUC) of 0.831 in DM, 0.883 in MIC, and 0.904 in the N group, compared with the SVM and KNN classifiers. <b>Conclusion:</b> MRI-based ML models have the potential to predict cognitive dysfunction in patients with T2DM. Compared with the SVM and KNN, the LR algorithm showed the best performance.
BibTeX
@inproceedings{zhigao2022mri,
  author = {XU Zhi-gao and Lili Zhao and Lei Yin and Yan Liu and Ying Ren and Guoqiang Yang and Jinlong Wu and Feng Long Gu and Xuesong Sun and Hui Yang and Taisong Peng and Jinfeng Hu and Xiaogeng Wang and Minghao Pang and Qiong Dai and Guojiang Zhang},
  title = {{MRI-based machine learning model: A potential modality for predicting cognitive dysfunction in patients with type 2 diabetes mellitus}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2022},
  doi = {10.3389/fbioe.2022.1082794},
}
Figure