Tada-DIP: Input-adaptive Deep Image Prior for One-shot 3D Image Reconstruction

2025-12-03

AI Summary

Deep Image Prior (DIP) suffers from overfitting and limited reconstruction quality in ill-posed 3D inverse problems. Method: We propose Tada-DIP—the first input-adaptive, unsupervised framework for 3D image reconstruction—requiring no training data. It integrates input-driven network initialization, gradient-domain denoising regularization, and an early-stopping optimization strategy to enable high-fidelity single-shot reconstruction. Contribution/Results: Tada-DIP introduces, for the first time, input adaptivity into the 3D DIP paradigm and explicitly regularizes the optimization trajectory to suppress overfitting. In sparse-view CT reconstruction, it significantly outperforms existing unsupervised baselines and matches the performance of fully supervised models trained on complete datasets. This establishes a highly efficient, robust, and data-free solution paradigm for ill-posed 3D inverse problems.

Technology Category

• Computer Vision: 3D Computer Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Deep Image Prior (DIP) has recently emerged as a promising one-shot neural-network based image reconstruction method. However, DIP has seen limited application to 3D image reconstruction problems. In this work, we introduce Tada-DIP, a highly effective and fully 3D DIP method for solving 3D inverse problems. By combining input-adaptation and denoising regularization, Tada-DIP produces high-quality 3D reconstructions while avoiding the overfitting phenomenon that is common in DIP. Experiments on sparse-view X-ray computed tomography reconstruction validate the effectiveness of the proposed method, demonstrating that Tada-DIP produces much better reconstructions than training-data-free baselines and achieves reconstruction performance on par with a supervised network trained using a large dataset with fully-sampled volumes.

Problem

• Develops a 3D image reconstruction method using deep image prior
• Addresses overfitting in one-shot neural network based reconstruction
• Improves reconstruction quality for sparse-view X-ray computed tomography

Innovation

• Input-adaptive Deep Image Prior for 3D reconstruction
• Combines input-adaptation and denoising regularization
• Avoids overfitting in one-shot 3D inverse problems