Data-Efficient Alignment in Medical Imaging via Reconfigurable Generative Networks

Divya Saxena, Jiannong Cao, Jiahao Xu, Tarun Kulshrestha

Abstract

Medical image-to-image (I2I) translation is a critical task for clinical diagnosis, yet it faces significant challenges due to anatomical misalignment and the scarcity of high-quality paired data. We present Reconfig-MIT, a framework designed to overcome these hurdles through two key innovations: a reconfigurable generative network that dynamically increases its width during training and a novel Recursive Cycle-Consistency Loss (R-CCL). By starting with a sparse architecture and expanding according to a cubic growth schedule, the model maintains a learning equilibrium between the generator and discriminator. Our results on MRI datasets demonstrate that Reconfig-MIT achieves superior alignment and translation quality even when trained on only 10% of available data, significantly outperforming standard fixed-architecture GANs.

Introduction

In medical imaging, translating between modalities—such as T1-weighted to T2-weighted MRI—is essential for comprehensive patient assessment. However, standard Generative Adversarial Networks (GANs) often fail in real-world scenarios due to "rigidity." Patient movement and physiological changes lead to spatial misalignments between source and target images, which fixed-width architectures struggle to resolve. Furthermore, the high cost and privacy concerns surrounding medical data result in small, sparse datasets that are prone to overfitting.

Existing solutions typically employ separate registration networks to align images before translation. However, these multi-stage pipelines can be computationally expensive and may fail to capture local, non-rigid deformations. Reconfig-MIT addresses these issues by making the generative model inherently robust to local inconsistencies through dynamic architectural adaptation.

Methods

Reconfigurable Generative Network

The core of Reconfig-MIT is a generator that evolves from a sparse state to a dense one. Instead of initializing a full-sized network, the model begins with a sparse configuration. During training, it dynamically adds neurons (increasing width) following a cubic growth function. This gradual expansion allows the network to first learn broad, low-level features before dedicating capacity to high-level anatomical complexities. This "Goldilocks" growth rate prevents the discriminator from overpowering the generator early in the process, ensuring a stable learning equilibrium.

The Reconfig-MIT training pipeline, showing the transition from a sparse network to a dense, reconfigurable generative network through dynamic width-wise growth.

Recursive Cycle-Consistency Loss (R-CCL)

To further enhance anatomical preservation, the framework introduces Recursive Cycle-Consistency Loss. While standard CycleGANs utilize a single "round-trip" (A → B → A), R-CCL implements an additional recursive loop ($x \to y' \to x' \to y'' \to x''$). This extended consistency check forces the model to be more disciplined in preserving spatial relationships and anatomical structures across multiple transformations, which is particularly effective in handling misaligned data.

Results and Discussion

The framework was evaluated using the BraTS 2018 dataset, consisting of T1 and T2 MRI brain scans. To simulate realistic clinical conditions, the authors introduced various levels of misalignment, including rigid rotations and non-rigid deformations. Crucially, experiments were conducted using only 10% of the training data.

Quantitative and Qualitative Performance

Reconfig-MIT demonstrated superior performance across all misalignment settings. In the most challenging scenarios (M.5), it achieved a PSNR of ~24.39, compared to ~23.50 for the baseline CycleGAN. Qualitative analysis via error maps revealed that Reconfig-MIT produces significantly fewer discrepancies, maintaining the integrity of delicate brain structures.

Qualitative comparison of T1-to-T2 translation results, error maps, and deformation fields (D.F) for CycleGAN, RegGAN, and Reconfig-MIT across different cases.

The sparse-to-dense growth strategy also ensures computational efficiency. By starting sparse, the model optimizes its parameter usage, often achieving better results with fewer total parameters than fixed-width counterparts.

Conclusion

Reconfig-MIT provides a robust and data-efficient framework for medical image translation and alignment. By combining dynamic network growth with recursive consistency losses, it eliminates the need for explicit pre-registration steps. While currently validated on 2D MRI slices, future work will explore extending this reconfigurable approach to 3D volumetric data and more diverse imaging modalities, such as CT-to-MRI translation.

Summary updated

Abstract

Recent advances in deep learning have witnessed many successful medical image translation models that learn correspondences between two visual domains. However, building robust mappings between domains is a significant challenge when handling misalignments caused by factors such as respiratory motion and anatomical changes. This issue is further exacerbated in scenarios with limited data availability, leading to a significant degradation in translation quality. In this paper, we introduce a novel data-efficient framework for aligning medical images via Reconfigurable Generative Network (Reconfig-MIT) for high-quality image translation. The key idea of Reconfig-MIT is to adaptively expand the generative network width within a Generative Adversarial Networks (GAN) architecture, initially expanding rapidly to capture low-level features and then slowing to refine high-level complexities. This dynamic network adaptation mechanism allows to adaptively learn at different rates, thus the model can better respond to deviations in the data caused by misalignments, while maintaining an effective equilibrium with the discriminator (D). We also introduce the Recursive Cycle-Consistency Loss (R-CCL), which extends the cycle consistency loss to effectively preserve key anatomical structures and their spatial relationships, improving translation quality. Extensive experiments show that Reconfig-MIT is a generic framework that enables easy integration with existing image translation methods, including those incorporating registration networks used for correcting misalignments, and provides robust and high-quality translation on paired and unpaired misaligned data in both data-rich and data-limited scenarios. https://github.com/IntellicentAI-Lab/Reconfig-MIT.
BibTeX
@inproceedings{saxena2025data,
  author = {Divya Saxena and Jiannong Cao and Jiahao Xu and Tarun Kulshrestha},
  title = {{Data-Efficient Alignment in Medical Imaging via Reconfigurable Generative Networks}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
  doi = {10.1109/wacv61041.2025.00719},
}
Figure