The Age-specific Alzheimer 's Disease Prediction with Characteristic Constraints in Nonuniform Time Span

TL;DR

The paper tackles predicting Alzheimer’s disease progression from irregularly spaced brain imaging by introducing T-GAN, a temporal GAN that jointly predicts future MRI/PET images and clinical indicators. The generator uses age-conditioned cross-attention to fuse age and image features, while two discriminators enforce image realism and preservation of disease-related indicators; a dynamic indicator loss handles missing data. Key contributions include an age-scaled pixel loss to balance long- and short-term sequences and an indicator discriminator that ties image synthesis to quantitative clinical metrics, enabling accurate long-term predictions with SSIM reaching up to $0.882$ and robust disease-feature fidelity. Across ablations and multi-modal validation, T-GAN outperforms baselines in preserving pathology features and producing high-quality longitudinal predictions, supporting potential clinical utility for early AD assessment and progression forecasting.

Abstract

Alzheimer's disease is a debilitating disorder marked by a decline in cognitive function. Timely identification of the disease is essential for the development of personalized treatment strategies that aim to mitigate its progression. The application of generated images for the prediction of Alzheimer's disease poses challenges, particularly in accurately representing the disease's characteristics when input sequences are captured at irregular time intervals. This study presents an innovative methodology for sequential image generation, guided by quantitative metrics, to maintain the essential features indicative of disease progression. Furthermore, an age-scaling factor is integrated into the process to produce age-specific MRI images, facilitating the prediction of advanced stages of the disease. The results obtained from the ablation study suggest that the inclusion of quantitative metrics significantly improves the accuracy of MRI image synthesis. Furthermore, the application of age-scaled pixel loss contributed to the enhanced iterative generation of MRI images. In terms of long-term disease prognosis, the Structural Similarity Index reached a peak value of 0.882, indicating a substantial degree of similarity in the synthesized images.

Figures (9)

• Figure 1: Simplified diagram of the model's structure. (a) Generator: The generator encodes and fuses the input images and ages features, decoding them into synthetic images. (b) Discriminator: The generated and real images are fed into the discriminator to obtain the quantitative indicators and image discriminator.
• Figure 2: Age Encoding Method. The age feature is encoded as a binary vector of length 1000, with a maximum age of 100.0 years, as the maximum age of samples in the dataset is 97.3. In this case, the age of the $i$th-period for the sample is 67.5 years, where the first 675 elements of the encoding vector are set to 1, and the remaining 325 elements are set to 0. Similarly, the encoding method can be applied to the $j$th-period. Subtracting the encoding vectors of two periods obtains the age difference vector$x_a$.
• Figure 3: The generator architecture diagram illustrates the utilization of the U-Net as the foundational network to construct the generator. (a) Image Encoder: Four downsampling convolutional layers extract pertinent features from the input image. (b) Condition Encoder: A multilayer perceptron is a conditional encoder for age-conditional information. (c) Feature Fusion Attention: The amalgamation of features is executed by the attention mechanism during the feature fusion process.(d) Output Decoder: The decoder module comprises four upsampling deconvolutional layers responsible for the final image's reconstruction. Ultimately, the resultant image is acquired by subjecting the output through a convolutional layer, succeeded by a tanh activation function.
• Figure 4: Simplified diagram of the model's discriminator. (a) Indicator Discriminator: The adversarial discriminator consists of five convolutional layers and outputs the authenticity of each pixel value in the image. (b) Adversarial Discriminator: The Adversarial Discriminator using the Feature Pyramid Network produces precise quantitative indicators by calculating the average of outputs from each scale in the network.
• Figure 5: Image Preprocessing Procedure