CLASSIFYING FRONTOTEMPORAL DEMENTIA SUBTYPES USING MRI AND DEEP TRANSFER LEARNING

MRI and deep learning classification models have been shown to reliably predict dementia. However, evidence regarding the utility of such models in frontotemporal dementia (FTD) is scarce, in part due to a dearth of large and accessible datasets. In this study, we employ transfer learning methods to overcome such data limitations.
We acquired data for 1,815 participants from the AD neuroimaging initiative (ADNI; n = 1,219), FRONTIER (n = 348), and frontotemporal lobar degeneration neuroimaging initiative (NIFD; n = 248). We used 3D grey-matter volumes as input in a DenseNet-121 model to classify AD using ADNI data. Then, we finetuned the AD model on the FRONTIER dataset to perform multiclass FTD classification. Both model iterations were validated using external datasets and included class-activation-mapping (CAM) techniques to identify key predictive brain regions.
The AD model had a binary classification accuracy of 90.54% and the FTD model a multiclass classification accuracy of 75.24%. These results remained consistent with external validation resulting in 89% and 77.34% accuracy, respectively. The FTD model had 99.2% accuracy for identifying controls, with accuracy varying across the FTD subtypes (semantic dementia: 91.4%; behavioural-variant FTD: 67.3%; progressive non-fluent aphasia: 5.9%). CAM results identified the cerebellum as key for detecting controls, frontostriatal regions for bvFTD, and temporal and parietal lobes for SD.
Overall, this study reveals the potential of transfer learning to maximise existing FTD datasets. These methods outperformed leading models in the literature; however, more work is required to refine the detection of PNFA and develop clinically accurate models.