Deformation-aware Temporal Generation for Early Prediction of Alzheimers Disease

TL;DR

This work addresses early Alzheimer's prediction from longitudinal brain MRIs with irregular temporal sampling. It introduces DATGN, a deformation-aware framework with a temporal interpolation module (E, B, P) and a deformation-guided temporal prediction module (spatial encoder, DT-Module, DT-LSTM, decoder) to generate future MRI sequences. The model achieves competitive image-quality metrics and, when augmented data is used for classification, significantly improves AD vs. NC and AD vs. MCI/NC accuracies, while qualitative results show atrophy-consistent evolution. Overall, DATGN enables improved early prediction by learning morphological changes over time through bidirectional deformation fields and self-supervised training, with promising clinical implications for timely intervention.

Abstract

Alzheimer's disease (AD), a degenerative brain condition, can benefit from early prediction to slow its progression. As the disease progresses, patients typically undergo brain atrophy. Current prediction methods for Alzheimers disease largely involve analyzing morphological changes in brain images through manual feature extraction. This paper proposes a novel method, the Deformation-Aware Temporal Generative Network (DATGN), to automate the learning of morphological changes in brain images about disease progression for early prediction. Given the common occurrence of missing data in the temporal sequences of MRI images, DATGN initially interpolates incomplete sequences. Subsequently, a bidirectional temporal deformation-aware module guides the network in generating future MRI images that adhere to the disease's progression, facilitating early prediction of Alzheimer's disease. DATGN was tested for the generation of temporal sequences of future MRI images using the ADNI dataset, and the experimental results are competitive in terms of PSNR and MMSE image quality metrics. Furthermore, when DATGN-generated synthetic data was integrated into the SVM vs. CNN vs. 3DCNN-based classification methods, significant improvements were achieved from 6. 21\% to 16\% in AD vs. NC classification accuracy and from 7. 34\% to 21. 25\% in AD vs. MCI vs. NC classification accuracy. The qualitative visualization results indicate that DATGN produces MRI images consistent with the brain atrophy trend in Alzheimer's disease, enabling early disease prediction.

Figures (6)

• Figure 1: The prediction of the progression of Alzheimer's disease involves generating future brain image sequences based on past brain image sequences. (a) represents the temporal sequence of past MRI brain images of a patient with missing data. In subfigure (b), the left part represents the temporal sequence after the missing data has been generated through an interpolation method, and the right part represents the future time-point sequence generated through a prediction method.
• Figure 2: DATGN consists of two modules: the temporal interpolation module and the temporal prediction module. The temporal interpolation module contains a deformation field estimation network and a interpolation network for completing the missing data; The temporal prediction module contains a prediction network for generating the temporal sequence of future brain images.
• Figure 3: The temporal interpolation module contains a deformation estimation network E, a interpolation network P and loss functions, wherein the deformation field estimation network $E$ estimate the inter-frame deformation fields ${I_{t{\rightarrow }t-1}^2}$,${I_{t{\rightarrow }t+1}^2}$ from the temporal brain images inputs ${I_{t-1}^1}$, ${I_{t+1}^1}$ ; and backward warping function B combines the original images ${I_{t-1}^1}$,${I_{t+1}^1}$ and the inter-frame deformation fields ${I_{t{\rightarrow }t-1}^2}$,${I_{t{\rightarrow }t+1}^2}$ to estimated bi-direction images $I_{t-1 \rightarrow t}^{3}$ ,$I_{t+1 \rightarrow t}^{3}$; the interpolation network P combines $I_{t-1 \rightarrow t}^{3}$,$I_{t+1 \rightarrow t}^{3}$ to generate the predicted intermediate frame $I_{t-1}^{pred}$, $I_{t+1}^{pred}$;${L^{Interpolation}}$ is global loss function,combine by $L^{B}$,$L^{P}$ and $L^{fusion}$, $L^{B}$ minimizing the difference between the true frame ${{I}_{t}^{gt}}$ with $I_{t-1 \rightarrow t}^{3}$,$I_{t+1 \rightarrow t}^{3}$,$L^{P}$ minimizing the difference between the true frame ${{I}_{t}^{gt}}$ with $I_{t}^{pred}$,$L^{fusion}$ minimizing the difference between the true frame ${{I}_{t}^{gt}}$ with the combination of $I_{t-1 \rightarrow t}^{3}$ and $I_{t}^{pred}$,$I_{t+1 \rightarrow t}^{3}$ and $I_{t}^{pred}$.
• Figure 4: The temporal prediction module includes a prediction network, a DT-Module and a DT-LSTM. The gray part represents the prediction network, the yellow part represents DT-Module and the blue part represents DT-LSTM. Among them, the prediction network receives the temporal sequence of brain images $X_{1:n}$ and the temporal sequence of deformation fields $D_{1:n}$, and outputs the temporal sequence of future brain images $X_{n+1:2n}$ after performing temporal modeling. The DT-Module with DT-LSTM are used for temporal modeling.
• Figure 5: A histogram of the distribution of data from 1872 subjects from the ADNI dataset, where the vertical axis represents the time span in months and the horizontal axis represents the sample size.