Radiomics Model for Frontotemporal Dementia Diagnosis Using T1-Weighted MRI

Abstract

Radiomics has been proposed as a useful approach to extrapolate novel morphological and textural information from brain Magnetic resonance images (MRI). Radiomics analysis has shown unique potential in the diagnostic work-up and in the follow-up of patients suffering from neurodegenerative diseases. However, the potentiality of this technique in distinguishing frontotemporal dementia (FTD) subtypes has so far not been investigated. In this study, we explored the usefulness of radiomic features in differentiating FTD subtypes, namely, the behavioral variant of FTD (bvFTD), the non-fluent and/or agrammatic (PNFA) and semantic (svPPA) variants of a primary progressive aphasia (PPA). Classification analyses were performed on 3 Tesla T1-weighted images obtained from the Frontotemporal Lobar Degeneration Neuroimaging Initiative. We included 49 patients with bvFTD, 25 patients with PNFA, 34 patients with svPPA, and 60 healthy controls. Texture analyses were conducted to define the first-order statistic and textural features in cortical and subcortical brain regions. Recursive feature elimination was used to select the radiomics signature for each pairwise comparison followed by a classification framework based on a support vector machine. Finally, 10-fold cross-validation was used to assess classification performances. The radiomics-based approach successfully identified the brain regions typically involved in each FTD subtype, achieving a mean accuracy of more than 80% in distinguishing between patient groups. Note mentioning is that radiomics features extracted in the left temporal regions allowed achieving an accuracy of 91 and 94% in distinguishing patients with svPPA from those with PNFA and bvFTD, respectively. Radiomics features show excellent classification performances in distinguishing FTD subtypes, supporting the clinical usefulness of this approach in the diagnostic work-up of FTD.

Behavioral variant frontotemporal dementia in patients with primary psychiatric disorder: A magnetic resonance imaging study

“…In the current study, we found that patients with PPD_bvFTD+ Di Benedetto et al, 2022;Meyer et al, 2017;Möller et al, 2016;Tafuri et al, 2022). They have generally reported optimal performance of this approach in distinguishing patients with bvFTD from controls.…”

“…Indeed, our findings provided new evidence about morphometric changes associated with bvFTD in patients with a long history of PPD. Different studies have applied machine learning‐based classification of patients with bvFTD based on patterns arising from MRI (Di Benedetto et al., 2022 ; Meyer et al., 2017 ; Möller et al., 2016 ; Tafuri et al., 2022 ). They have generally reported optimal performance of this approach in distinguishing patients with bvFTD from controls.…”

Explainable machine learning radiomics model for Primary Progressive Aphasia classification

“…Regarding the differentiation between the two PPA variants, our combined model achieved a diagnostic accuracy of 93.7% on the test set. This result overcomes the state-of-the-art performances achieved using only classical morphometry measurements ( Agosta et al, 2015 ; Bisenius et al, 2017 ; Kim et al, 2019 ; Lampe et al, 2022 ) and radiomics on gray matter ROIs ( Tafuri et al, 2022a ), in conjunction with machine learning systems like Support Vector Machine, Random Forest, or Linear Discriminant Analysis. It’s worth noting that, contrary to the comparison between pathological and healthy subjects, clinical/cognitive variables were unable to correctly identify the PPA phenotype achieving a suboptimal accuracy of classification of the 77.1%.…”

“…Consequently, numerous studies have employed the radiomics approach to uncover imaging biomarkers in cancers ( Vial et al, 2018 ) and, more recently, to evaluate diagnosis and prognosis in other diseases, including neurodegenerative conditions ( Salvatore et al, 2019 ; Feng and Ding, 2020 ; Tafuri et al, 2022a ). In particular, classification models have been developed by extracting high-dimensional sets of radiomics measures in specific brain regions and then combining feature selectors and machine learning algorithms to distinguish between diagnostic categories ( Feng et al, 2018 ; Ranjbar et al, 2019 ; Tafuri et al, 2022a , b ; Rajagopalan et al, 2023 ). Despite the optimal performance obtained by these classification frameworks, however, the estimation of each feature contribution to the model’s classification is often unclear limiting the interpretability of the results.…”

Transdiagnostic brain module dysfunctions across sub-types of frontotemporal dementia: a connectome-based investigation

“…Machine learning, tailored for intricate multivariate data, holds promise as an adjunct in the diagnostic process by offering decision support [25]. Previous studies have use machine learning algorithms to classify between the subtypes of dementia [25][26][27] and subtypes of FTD [28,29] in structural MRI. The accuracy of machine learning model between subtypes of FTD was ranged from 0.7-0.94.…”