Disentangled Representation Learning for Capturing Individualized Atrophy of Alzheimer's Disease

1409 

Abstract Submission 

Zhuangzhuang Li¹, Yibao Sun¹, Kun Zhao¹, Yong Liu¹

¹School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China

Zhuangzhuang Li  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Yibao Sun  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Kun Zhao  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Yong Liu  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Alzheimer's disease (AD) is a highly heterogeneous neurodegenerative disease, characterized by cognitive decline, irreversible memory loss, psychiatric symptoms, and brain atrophy. Capturing individualized pathological changes from healthy controls (HC) to AD is critical for early diagnosis and targeted treatment. However, existing studies focus on the group averages, assuming homogeneity between AD patients. We proposed a novel deep disentangled generative model (DDGM) for capturing individualized neuroanatomical alterations. The proposed method disentangles AD into "realistic" healthy counterfactual images and residual maps. The residual map can localize hypothetical abnormalities within a normal brain image that may cause it to be diagnosed with AD.

DDGM is composed of two encoder-decoder branches and a decoder branch. One encoder-decoder branch generates pseudo-healthy images by adversarial training, and the other is for synthetic disease residual maps. The decoder branch is for facilitating the training process. From Fig.1B, we can see that an input image can synthesize a realistic pseudo-healthy image, reconstruct the input image, and produce a residual map. The residual map indicates the underlying abnormal changes with disease progression.

Preserving the healthiness and subject identity of pseudo-healthy images for medical image translation tasks is important. The biological validity of synthetic pseudo-healthy images was evaluated by longitudinal data (HC convert to AD) from the ADNI database. We calculated the structural similarity index (SSIM) between synthetic pseudo-healthy images and longitudinal HC images to guarantee healthiness when AD images are fed into DDGM. The generated pseudo-healthy images and longitudinal HC images have a similar image structure (SSIM: 89.98% ± 0.04, Fig. 1C). To preserve the subject identity, we computed the SSIM values between synthetic pseudo-healthy images and original images across the testing subjects. The mean SSIM values between synthetic HC images produced by AD and longitudinal HC images are 98.04% ± 0.13. The generated HC images and original HC images have a similar image structure (SSIM: 98.71% ± 0.20, Fig. 1D) when HC images are fed into DDGM. The residual maps generated by DDGM can reflect the brain atrophy patterns of AD individuals. Finally, we used the ADNI dataset to explore the brain atrophy patterns of each condition in the human brain, such as HC, PMCI, SMCI, and AD. As shown in Fig. 1E, it is observed that AD, PMCI, SMCI, and HC showed an increasing trend of brain atrophy. The regions with the most severe gray matter atrophy were found in the hippocampus, amygdala, and part of the temporal lobe regions by AD abnormal residual maps.

We proposed a novel DRL-based model for capturing individualized neuroanatomical alterations. The proposed model can synthesize realistic healthy counterfactual images and produce saliency residual maps to indicate the underlying abnormal regions for interpretation. The saliency residual maps can help neurologists to understand the changes in disease progression. We believe that our proposed method will open new avenues for improving individualized diagnosis and the development of precision medicine for clinical intervention.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling ¹

Other - Alzheimer's disease, individualized pathological changes, deep disentangled generative model