Tortuosity and discrete compactness biomarkers for machine learning-based classification of mild cognitive impairment

Abstract

Objective

This study aimed to assess the effectiveness of tortuosity and discrete compactness metrics in analyzing the amygdala’s morphology to differentiate healthy control individuals from patients diagnosed with mild cognitive impairment using magnetic resonance imaging.

Methods

The analysis included a total of 74 participants, comprising 37 healthy control subjects and 37 mild cognitive impairment patients. Imaging data were sourced from the ADNI database. The amygdala regions (both hemispheres) were segmented, and measurements for volume, normalized volume, discrete compactness, and tortuosity were computed. Statistical tests and automatic classifier training of Support Vector Machines, K-nearest Neighbors, Randon Forest and Artificial Neural Network were conducted to identify significant group differences. The machine learning algorithms were trained with the proposed metrics with a partition of 60–40 subjects for training and testing. The training consisted of hyperparameter optimization with a 5-fold cross validation.

Results

The statistical analysis revealed significant differences (p < 0.01) across all evaluated metrics, with the most pronounced alterations observed in discrete compactness and tortuosity within the right hemisphere.

The application of the previously described algorithms demonstrated that the proposed biomarkers—tortuosity and discrete compactness—offered greater discriminative power compared to traditional volume-based measures. When incorporated into the classification models, these features enhanced performance, yielding a test accuracy of 82.14 %, area under the curve values between 88.27 % and 91.33 %, and F-scores ranging from 81.48 % to 83.87 %. These findings underscore the potential of tortuosity and discrete compactness as sensitive and robust imaging biomarkers for the early detection of mild cognitive impairment.

Conclusions

These findings demonstrate that tortuosity and discrete compactness are more sensitive than conventional volume-based metrics in capturing morphological alterations of the amygdala in mild cognitive impairment. When integrated into machine learning models—Support Vector Machines, K-nearest Neighbors, Random Forest, and Artificial Neural Networks—these features enhanced classification performance, achieving a test accuracy of 82.14 %, area under the curve values between 88.27 % and 91.33 %, and F-scores ranging from 81.48 % to 83.87 %.

Significance

The results suggest that tortuosity and discrete compactness may serve as robust and informative imaging biomarkers for the early detection of mild cognitive impairment. Their ability to outperform traditional morphological metrics in both statistical discrimination and machine learning classification highlights their potential for clinical application in computer-aided diagnosis systems.