Cas-DiffCom: Cascaded Diffusion Model for Infant Longitudinal Super-Resolution 3D Medical Image Completion

Lianghu Guo, Tianli Tao, Xinyi Cai, Zihao Zhu, Jiawei Huang, Lixuan Zhu, Zhuoyang Gu, Haifeng Tang, Rui Zhou, Siyan Han, Yan Liang, Qing Yang, Dinggang Shen, Han Zhang

Abstract

Longitudinal MRI studies are essential for understanding early neurodevelopment, yet they are frequently plagued by missing data due to subject attrition or motion artifacts. We present Cas-DiffCom, a novel two-stage diffusion-based framework designed to complete missing 3D brain scans. By combining a multimodal generative stage with a high-fidelity super-resolution stage, Cas-DiffCom produces synthetic MRIs that are anatomically accurate and consistent with an individual's unique developmental trajectory. Evaluated on the Baby Connectome Project (BCP) dataset, our model significantly outperforms existing GAN-based methods in both image quality and downstream scientific utility.

Introduction

The first two years of life involve rapid neurodevelopmental changes. Researchers typically use longitudinal MRI—scanning the same infant at multiple time points—to track these changes. However, acquiring complete longitudinal datasets is challenging; infants are difficult subjects, often leading to missing time points in a child's developmental record.

Existing AI methods, such as Generative Adversarial Networks (GANs), often struggle with this task, frequently suffering from "model collapse" (producing average-looking brains) or failing to maintain subject-specific consistency across different ages. Cas-DiffCom addresses these issues by generating high-resolution, subject-specific missing scans realistic enough for clinical research.

Methodology

Cas-DiffCom (Cascaded Diffusion Model for Completion) decomposes the completion problem into two distinct stages to manage the computational complexity and stability of generating high-resolution 3D volumes.

1. Low-Resolution Structure Generation

The first stage generates a structured, low-resolution (80x96x80) 3D image. This stage utilizes a novel Asynchronous Multimodal (AsMM) block to process two distinct inputs: the target age and existing MRI scans from other time points. Unlike traditional concatenation methods, the AsMM block disentangles these inputs. It uses an attention mechanism for age-based global context and skip-connections for structural image features, ensuring the model captures how a specific brain should appear at a given developmental milestone.

2. High-Resolution Refinement

The second stage takes the low-resolution output and performs 3D super-resolution, upsampling the volume to 160x192x160. This stage adds the fine-grained textures and sharpness necessary for medical analysis. The framework is built on Denoising Diffusion Probabilistic Models (DDPM) and utilizes Denoising Diffusion Implicit Models (DDIM) sampling to accelerate inference.

Overview of the Cas-DiffCom architecture, including the AsMM block and Super-resolution block details.

Experimental Setup

The model was trained and evaluated using the Baby Connectome Project (BCP) dataset, which includes 654 T1-weighted MRI scans from 340 subjects ranging from 2 weeks to 26 months of age. A test set of 30 MRI scans was used to evaluate performance across the entire developmental spectrum.

Results and Discussion

Cas-DiffCom significantly outperformed standard baselines in both quantitative metrics and qualitative fidelity. The model achieved a PSNR of 24.15 and an SSIM of 0.81, compared to 17.76 and 0.72 for conditional GAN (cGAN) baselines.

Qualitatively, the generated images exhibit clear contrast between white and grey matter, a feature typically difficult to capture in infants due to ongoing myelination. Furthermore, growth trajectories for grey matter, white matter, and cerebrospinal fluid (CSF) derived from the synthetic images closely matched the ground truth distributions of the real dataset, confirming the model's scientific validity.

Visual results demonstrating the longitudinal consistency of generated T1-weighted images and segmentations from 3 to 26 months.

Conclusion

Cas-DiffCom demonstrates the power of cascaded diffusion models for high-fidelity medical image completion. By disentangling age and structural guidance, the model maintains individual developmental trajectories better than previous approaches. Future work will focus on reducing the computational cost of 3D diffusion sampling and testing the model's generalizability across different MRI modalities (e.g., T2-weighted) and multi-site data.

Summary updated

BibTeX
@inproceedings{guo2024cas,
  author = {Lianghu Guo and Tianli Tao and Xinyi Cai and Zihao Zhu and Jiawei Huang and Lixuan Zhu and Zhuoyang Gu and Haifeng Tang and Rui Zhou and Siyan Han and Yan Liang and Qing Yang and Dinggang Shen and Han Zhang},
  title = {{Cas-DiffCom: Cascaded Diffusion Model for Infant Longitudinal Super-Resolution 3D Medical Image Completion}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635663},
}
Figure