Spinal and Bulbar Muscular Atrophy is a rare X-linked neurodegenerative disorder primarily affecting lower motor neurons. Although white-matter alterations have been documented, evidence concerning cortical gray-matter involvement remains limited and inconsistent. This thesis investigates whether structural brain differences between patients with SBMA and healthy controls can be identified using complementary supervised and unsupervised machine-learning approaches. A dataset of 37 T1-weighted structural MRI scans, comprising 20 patients with SBMA and 17 healthy controls, was analyzed using two complementary approaches. In the supervised analysis, GM was extracted using a 1,000-region cortical parcellation and two regional representations were considered: raw regional GM and GM normalized by total GM volume. The ten most informative regions were selected using mutual information within a Leave-One-Out Cross-Validation framework. Four classification algorithms—Support Vector Machine, Logistic Regression, Random Forest, and XGBoost—were evaluated. Explainable artificial intelligence based on SHAP values was subsequently used to characterize the contribution and direction of individual brain regions to the model predictions. In parallel, an unsupervised community detection analysis was performed by constructing subject-specific graphs in which brain parcels were represented as nodes and spatially connected parcels were weighted according to GM similarity. Separate multiplex networks were constructed for the SBMA and healthy control cohorts, and the Leiden algorithm with consensus clustering was used to identify stable communities. Classification performance was substantially higher when proportional regional GM measures were used compared with raw GM measures. L1-regularized Logistic Regression achieved the best overall balance between sensitivity and specificity, with an accuracy of 0.84 and an F1-score of 0.84. SHAP analyses consistently identified regions of the frontal operculum–insula within the salience/ventral attention network and prefrontal regions of the default mode network as important contributors to the classification. Lower proportional GM values in these regions were generally associated with predictions of SBMA. The community detection analysis revealed broadly similar organizational structures between the two cohorts, with eight communities identified in healthy controls and seven in SBMA. The resulting partitions showed high agreement, with an Adjusted Rand Index of 0.962 and a Normalized Mutual Information of 0.960. Differences were concentrated in a relatively small number of regions, primarily within the left somatosensory motor network and, to a lesser extent, the dorsal attention network. Overall, these findings demonstrate the potential of machine learning and network-based approaches to investigate brain involvement in SBMA. The results suggest that localized patterns of regional GM may contain information capable of distinguishing SBMA from healthy controls, including regions outside the primary motor system, while the overall organization of GM similarity remains largely preserved. Given the limited sample size, the identified regions should be validated in larger, independent datasets.

Spinal and Bulbar Muscular Atrophy is a rare X-linked neurodegenerative disorder primarily affecting lower motor neurons. Although white-matter alterations have been documented, evidence concerning cortical gray-matter involvement remains limited and inconsistent. This thesis investigates whether structural brain differences between patients with SBMA and healthy controls can be identified using complementary supervised and unsupervised machine-learning approaches. A dataset of 37 T1-weighted structural MRI scans, comprising 20 patients with SBMA and 17 healthy controls, was analyzed using two complementary approaches. In the supervised analysis, GM was extracted using a 1,000-region cortical parcellation and two regional representations were considered: raw regional GM and GM normalized by total GM volume. The ten most informative regions were selected using mutual information within a Leave-One-Out Cross-Validation framework. Four classification algorithms—Support Vector Machine, Logistic Regression, Random Forest, and XGBoost—were evaluated. Explainable artificial intelligence based on SHAP values was subsequently used to characterize the contribution and direction of individual brain regions to the model predictions. In parallel, an unsupervised community detection analysis was performed by constructing subject-specific graphs in which brain parcels were represented as nodes and spatially connected parcels were weighted according to GM similarity. Separate multiplex networks were constructed for the SBMA and healthy control cohorts, and the Leiden algorithm with consensus clustering was used to identify stable communities. Classification performance was substantially higher when proportional regional GM measures were used compared with raw GM measures. L1-regularized Logistic Regression achieved the best overall balance between sensitivity and specificity, with an accuracy of 0.84 and an F1-score of 0.84. SHAP analyses consistently identified regions of the frontal operculum–insula within the salience/ventral attention network and prefrontal regions of the default mode network as important contributors to the classification. Lower proportional GM values in these regions were generally associated with predictions of SBMA. The community detection analysis revealed broadly similar organizational structures between the two cohorts, with eight communities identified in healthy controls and seven in SBMA. The resulting partitions showed high agreement, with an Adjusted Rand Index of 0.962 and a Normalized Mutual Information of 0.960. Differences were concentrated in a relatively small number of regions, primarily within the left somatosensory motor network and, to a lesser extent, the dorsal attention network. Overall, these findings demonstrate the potential of machine learning and network-based approaches to investigate brain involvement in SBMA. The results suggest that localized patterns of regional GM may contain information capable of distinguishing SBMA from healthy controls, including regions outside the primary motor system, while the overall organization of GM similarity remains largely preserved. Given the limited sample size, the identified regions should be validated in larger, independent datasets.

Machine Learning for Brain MRI Analysis in Spinal and Bulbar Muscular Atrophy

BONIN, GIORGIA
2025/2026

Abstract

Spinal and Bulbar Muscular Atrophy is a rare X-linked neurodegenerative disorder primarily affecting lower motor neurons. Although white-matter alterations have been documented, evidence concerning cortical gray-matter involvement remains limited and inconsistent. This thesis investigates whether structural brain differences between patients with SBMA and healthy controls can be identified using complementary supervised and unsupervised machine-learning approaches. A dataset of 37 T1-weighted structural MRI scans, comprising 20 patients with SBMA and 17 healthy controls, was analyzed using two complementary approaches. In the supervised analysis, GM was extracted using a 1,000-region cortical parcellation and two regional representations were considered: raw regional GM and GM normalized by total GM volume. The ten most informative regions were selected using mutual information within a Leave-One-Out Cross-Validation framework. Four classification algorithms—Support Vector Machine, Logistic Regression, Random Forest, and XGBoost—were evaluated. Explainable artificial intelligence based on SHAP values was subsequently used to characterize the contribution and direction of individual brain regions to the model predictions. In parallel, an unsupervised community detection analysis was performed by constructing subject-specific graphs in which brain parcels were represented as nodes and spatially connected parcels were weighted according to GM similarity. Separate multiplex networks were constructed for the SBMA and healthy control cohorts, and the Leiden algorithm with consensus clustering was used to identify stable communities. Classification performance was substantially higher when proportional regional GM measures were used compared with raw GM measures. L1-regularized Logistic Regression achieved the best overall balance between sensitivity and specificity, with an accuracy of 0.84 and an F1-score of 0.84. SHAP analyses consistently identified regions of the frontal operculum–insula within the salience/ventral attention network and prefrontal regions of the default mode network as important contributors to the classification. Lower proportional GM values in these regions were generally associated with predictions of SBMA. The community detection analysis revealed broadly similar organizational structures between the two cohorts, with eight communities identified in healthy controls and seven in SBMA. The resulting partitions showed high agreement, with an Adjusted Rand Index of 0.962 and a Normalized Mutual Information of 0.960. Differences were concentrated in a relatively small number of regions, primarily within the left somatosensory motor network and, to a lesser extent, the dorsal attention network. Overall, these findings demonstrate the potential of machine learning and network-based approaches to investigate brain involvement in SBMA. The results suggest that localized patterns of regional GM may contain information capable of distinguishing SBMA from healthy controls, including regions outside the primary motor system, while the overall organization of GM similarity remains largely preserved. Given the limited sample size, the identified regions should be validated in larger, independent datasets.
2025
Machine Learning for Brain MRI Analysis in Spinal and Bulbar Muscular Atrophy
Spinal and Bulbar Muscular Atrophy is a rare X-linked neurodegenerative disorder primarily affecting lower motor neurons. Although white-matter alterations have been documented, evidence concerning cortical gray-matter involvement remains limited and inconsistent. This thesis investigates whether structural brain differences between patients with SBMA and healthy controls can be identified using complementary supervised and unsupervised machine-learning approaches. A dataset of 37 T1-weighted structural MRI scans, comprising 20 patients with SBMA and 17 healthy controls, was analyzed using two complementary approaches. In the supervised analysis, GM was extracted using a 1,000-region cortical parcellation and two regional representations were considered: raw regional GM and GM normalized by total GM volume. The ten most informative regions were selected using mutual information within a Leave-One-Out Cross-Validation framework. Four classification algorithms—Support Vector Machine, Logistic Regression, Random Forest, and XGBoost—were evaluated. Explainable artificial intelligence based on SHAP values was subsequently used to characterize the contribution and direction of individual brain regions to the model predictions. In parallel, an unsupervised community detection analysis was performed by constructing subject-specific graphs in which brain parcels were represented as nodes and spatially connected parcels were weighted according to GM similarity. Separate multiplex networks were constructed for the SBMA and healthy control cohorts, and the Leiden algorithm with consensus clustering was used to identify stable communities. Classification performance was substantially higher when proportional regional GM measures were used compared with raw GM measures. L1-regularized Logistic Regression achieved the best overall balance between sensitivity and specificity, with an accuracy of 0.84 and an F1-score of 0.84. SHAP analyses consistently identified regions of the frontal operculum–insula within the salience/ventral attention network and prefrontal regions of the default mode network as important contributors to the classification. Lower proportional GM values in these regions were generally associated with predictions of SBMA. The community detection analysis revealed broadly similar organizational structures between the two cohorts, with eight communities identified in healthy controls and seven in SBMA. The resulting partitions showed high agreement, with an Adjusted Rand Index of 0.962 and a Normalized Mutual Information of 0.960. Differences were concentrated in a relatively small number of regions, primarily within the left somatosensory motor network and, to a lesser extent, the dorsal attention network. Overall, these findings demonstrate the potential of machine learning and network-based approaches to investigate brain involvement in SBMA. The results suggest that localized patterns of regional GM may contain information capable of distinguishing SBMA from healthy controls, including regions outside the primary motor system, while the overall organization of GM similarity remains largely preserved. Given the limited sample size, the identified regions should be validated in larger, independent datasets.
SBMA
MRI
Random Forest
Louvain
Brain Connectivity
File in questo prodotto:
File Dimensione Formato  
Bonin_Giorgia.pdf

accesso aperto

Dimensione 2.72 MB
Formato Adobe PDF
2.72 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/115822