Longitudinal White Matter Hyperintensity Segmentation with LSTM-Enhanced U-Net
Resumo
White matter hyperintensities (WMHs) are commonly accepted biomarkers of brain aging and neurodegeneration, typically observed on magnetic resonance imaging. Segmenting WMHs using U-Net-based convolutional neural networks (CNNs) is a viable and widely demonstrated technique to automate an otherwise laborious and tedious manual task. Although automation of cross-sectional WMH segmentation has been demonstrated, failure to appropriately process temporal information may decrease accuracy and limit understanding of lesion progression. Here, we investigate the combination of two convolutional long short-term memory (LSTM) approaches and U-Net models to segment WMHs in young and old adults. Using information acquired at two visits in 50 healthy individuals, we analyzed model-task performance using F-measure of three variants: 1) baseline standard U-Net, 2) Partial LSTM U-Net, and 3) Full LSTM U-Net. Results showed that while LSTM-based models outperform the baseline model in F-measure, notable differences emerged in how effectively they detected WMHs of varying size and distribution between age groups. These findings highlight the need to account for age-related heterogeneity in lesion morphology when designing and evaluating automated longitudinal WMH segmentation tools, and suggest that age- and size-aware modeling may lead to more robust predictions.
