Adaptive Independent Subspace Analysis of Brain Magnetic Resonance Imaging Data

Qiao Ke, Jiangshe Zhang, Wei Wei, Robertas Damasevicius, Marcin WozniakIEEE Access 2019 · DOI

Abstract

This paper presents a novel framework for the automated diagnosis of Autism Spectrum Disorder (ASD) using Adaptive Independent Subspace Analysis (AISA) on brain MRI data. By relaxing the independence constraints of traditional Independent Component Analysis (ICA), AISA effectively models coordinated brain activity within subspaces. Combined with multi-category texture analysis and k-Nearest Neighbor classification, the proposed method achieves a classification accuracy of 94.7% on pediatric MRI scans, demonstrating its potential as a robust diagnostic tool.

Introduction

Magnetic Resonance Imaging (MRI) provides a wealth of data about brain structure and function, but its sheer complexity makes manual clinical analysis time-consuming and prone to human error. While techniques like Independent Component Analysis (ICA) are often used to find meaningful patterns in this data, they have a major limitation: they assume all "sources" in the brain are statistically independent. In reality, brain regions often work in coordinated groups or "subspaces." This research addresses the need for a more flexible mathematical model that can account for these dependencies, specifically to help clinicians identify the subtle structural and functional markers of ASD in pediatric patients.

Methodology

The proposed framework utilizes a multi-stage pipeline designed to transform raw MRI pixels into high-confidence diagnostic predictions.

Preprocessing and AISA

Initially, MRI slices undergo resizing, sharpening, and normalization. A critical step is "bias field correction," which removes artifacts caused by magnetic field inhomogeneities. The core innovation of this work is the application of Adaptive Independent Subspace Analysis (AISA). Unlike standard ICA, AISA relaxes the independence constraint by grouping related components into subspaces and utilizing a "super-Gaussian" probability distribution. This approach better captures the sparse, non-linear nature of brain signals, acknowledging that while different functional systems might be independent, the neurons within those systems are highly correlated.

The multi-stage pipeline for ASD diagnosis using AISA and texture analysis.

Texture Analysis and Classification

Following subspace extraction, the researchers perform texture analysis on the AISA-derived subspaces. AISA acts as a powerful filter bank; the filters learned by the algorithm effectively act as edge detectors at various frequencies and orientations, highlighting relevant structural patterns. The team extracts 49 distinct features across five categories, including Gray-Level Co-occurrence Matrices (GLCM) and Neighborhood Gray-Tone Difference Matrices (NGTDM). These features describe the spatial relationship and "rhythm" of pixels, acting as significant markers for ASD that might be invisible to the naked eye. To manage the high dimensionality of the feature set, t-SNE is used for dimensionality reduction, followed by a k-Nearest Neighbor (k-NN) classifier for final diagnostic prediction.

Experimental Setup

The study utilized the NAMIC dataset, containing T1-weighted MRI brain scans from two autistic children and two control subjects, scanned at ages 2 and 4. Although the subject count is small, the analysis is performed on a per-slice basis. Each MRI file contains 192 grayscale slices (256x256 pixels), providing a substantial volume of image data for the AISA algorithm to learn robust features.

Results and Analysis

The AISA-based framework demonstrated high discriminative power in distinguishing between autistic and neurotypical scans. The key performance metrics include:

  • Accuracy: 94.7%
  • F-score: 0.9356
  • Specificity: 0.9482 (Ability to correctly identify "normal" scans)
  • Sensitivity: 0.9229 (Ability to correctly identify "autistic" scans)

Statistical analysis (ANOVA) confirmed that the texture features extracted from AISA subspaces were significantly different between the groups. Features such as "Local Homogeneity" and "Inverse Difference" yielded very low p-values (p < 0.001), validating their importance in the classification process.

Conclusion

This study highlights the effectiveness of modeling brain activity through independent subspaces rather than isolated components. While the high accuracy is promising, the primary limitation remains the small sample size of unique individuals. Future work will focus on expanding the dataset to validate clinical utility across the diverse autism spectrum and exploring additional feature sets, such as image moments and shape descriptors, to capture macro-structural changes in brain anatomy.

Summary updated

BibTeX
@article{ke2019adaptive,
  author = {Qiao Ke and Jiangshe Zhang and Wei Wei and Robertas Damasevicius and Marcin Wozniak},
  title = {{Adaptive Independent Subspace Analysis of Brain Magnetic Resonance Imaging Data}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2893496},
}
Figure