After a detailed analysis of the obtained characteristics and considering the current state of the art in the field, an approach based on the application of computational learning techniques has been adopted. This approach was preceded by a feature selection phase through a correlation study using the distance matrix based on the Pearson coefficient. Three distinct supervised learning methods were employed: Random Forest generated a ranking of features and was used to obtain classification results; Support Vector Machine (SVM) was used to compare and confirm classification results. Finally, the architecture of Convolutional Neural Networks (CNN) was applied in four distinct case studies, considering combinations of qMRI characteristics and radiomic features extracted from either the classical 3DT1 or novel approach T1/T2 images. The dataset was initially imbalanced, but it was balanced using resampling techniques, specifically SMOTE oversampling for the minority class, which is the Gd lesions class. After applying SMOTE, there were 626 lesions for class 0, 594 lesions for class 1, and a total of 1220 lesions. The dataset was subsequently divided into a training set (80%), a validation set (10%), and a test set (10%). RandomOverSampling was then exclusively used to further enrich the training set. The results obtained from the neural network analysis were carefully evaluated using both the validation and test sets, with a focus on accuracy and analysis of the confusion matrix. This evaluation confirmed the validity and reliability of the results obtained with 98.3% accuracy and a classification error of 3.17%, highlighting the relevance and effectiveness of the adopted methodological approach. In conclusion, the proposed methodological approach has proven to be effective in discriminating lesions as enhanced or non-enhanced with Gadolinium using computational learning and radiomics techniques applied to clinical and advanced MRI data. The use of predictive models based on supervised learning algorithms allowed for accurate classification of lesions without the need for contrast media. The results were validated through rigorous statistical analysis and evaluation on independent datasets, confirming the robustness and reliability of the proposed model. This study presents new possibilities for the diagnosis and monitoring of multiple sclerosis, offering a promising approach to reduce the use of Gadolinium and improve disease management.
Multiple sclerosis (MS) is an autoimmune inflammatory disease of the central nervous system characterized by a demyelination process that leads to loss of myelin and the formation of brain lesions. Currently, magnetic resonance imaging (MRI) plays a crucial role in diagnosing and monitoring disease activity. Specifically, the use of the contrast agent Gadolinium (Gd) is widely adopted to discriminate active or inflamed lesions for diagnostic and treatment purposes. However, the use of Gd is associated with potential toxicity risks and is inherently limited. This study aims to develop a predictive model for classifying lesions as enhanced or non-enhanced with Gadolinium starting from clinical and advanced MRI, in order to reduce the need for contrast media. The methodology is based on the analysis of a patient cohort obtained from the MSCentre XNAT database, which provides high-quality MRI data. In this context, various quantitative MRI (qMRI) metrics were considered and evaluated using statistical techniques, including boxplot analysis to identify any outliers and t-tests to assess statistical significance. Subsequently, a subset of 15 patients with brain lesions was identified, who presented both enhanced (1) and non-enhanced (0) lesions after administration of Gd. Overall, 626 class 0 lesions and 57 class 1 lesions were identified. This patient cohort represents a valuable resource for the analysis and characterization of brain lesions in correlation with radiomic and clinical characteristics. The Pyradiomics open-source package in Python was used to explore and apply radiomics, in order to extract features from the corresponding regions of interest (ROIs) of the identified lesions.
UN MODELLO DI RETE NEURALE PREDITTIVA PER LA CLASSIFICAZIONE DELLE LESIONI GADOLINIO-POTENZIATE NELLA SCLEROSI MULTIPLA
MALERBA, DAVIDE
2022/2023
Abstract
After a detailed analysis of the obtained characteristics and considering the current state of the art in the field, an approach based on the application of computational learning techniques has been adopted. This approach was preceded by a feature selection phase through a correlation study using the distance matrix based on the Pearson coefficient. Three distinct supervised learning methods were employed: Random Forest generated a ranking of features and was used to obtain classification results; Support Vector Machine (SVM) was used to compare and confirm classification results. Finally, the architecture of Convolutional Neural Networks (CNN) was applied in four distinct case studies, considering combinations of qMRI characteristics and radiomic features extracted from either the classical 3DT1 or novel approach T1/T2 images. The dataset was initially imbalanced, but it was balanced using resampling techniques, specifically SMOTE oversampling for the minority class, which is the Gd lesions class. After applying SMOTE, there were 626 lesions for class 0, 594 lesions for class 1, and a total of 1220 lesions. The dataset was subsequently divided into a training set (80%), a validation set (10%), and a test set (10%). RandomOverSampling was then exclusively used to further enrich the training set. The results obtained from the neural network analysis were carefully evaluated using both the validation and test sets, with a focus on accuracy and analysis of the confusion matrix. This evaluation confirmed the validity and reliability of the results obtained with 98.3% accuracy and a classification error of 3.17%, highlighting the relevance and effectiveness of the adopted methodological approach. In conclusion, the proposed methodological approach has proven to be effective in discriminating lesions as enhanced or non-enhanced with Gadolinium using computational learning and radiomics techniques applied to clinical and advanced MRI data. The use of predictive models based on supervised learning algorithms allowed for accurate classification of lesions without the need for contrast media. The results were validated through rigorous statistical analysis and evaluation on independent datasets, confirming the robustness and reliability of the proposed model. This study presents new possibilities for the diagnosis and monitoring of multiple sclerosis, offering a promising approach to reduce the use of Gadolinium and improve disease management.È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/17332