[Paper Review] An Asymetric Cycle-Consistency Loss for Dealing with Many-to-One Mappings in Image Translation: A Study on Thigh MR Scans

Michael Gadermayr, Maximilian Tschuchnig+5 more|arXiv (Cornell University)|Apr 23, 2020

Generative Adversarial Networks and Image Synthesis13 references4 citations

TL;DR

This paper proposes an asymmetric cycle-consistency loss to address many-to-one mappings in image-to-image translation for medical imaging, specifically thigh MRI scans with fat-infiltrated muscles. By modifying the cycle-consistency objective to account for multiple source images mapping to a single target, the method improves translation quality and segmentation performance without increasing model complexity.

ABSTRACT

Generative adversarial networks using a cycle-consistency loss facilitate unpaired training of image-translation models and thereby exhibit a very high potential in manifold medical applications. However, the fact that images in one domain potentially map to more than one image in another domain (e.g. in case of pathological changes) exhibits a major challenge for training the networks. In this work, we offer a solution to improve the training process in case of many-to-one mappings by modifying the cycle-consistency loss. We show formally and empirically that the proposed method improves the performance significantly without radically changing the architecture and without increasing the overall complexity. We evaluate our method on thigh MRI scans with the final goal of segmenting the muscle in fat-infiltrated patients' data.

Motivation & Objective

• To address the challenge of many-to-one mappings in unpaired image translation, common in medical imaging where multiple pathological scans may map to a single healthy reference.
• To improve training stability and translation quality in generative adversarial networks when mapping from diverse pathological images to a single normal reference domain.
• To maintain model simplicity and computational efficiency while enhancing performance on medical image translation tasks.
• To evaluate the method on thigh MRI scans for accurate muscle segmentation in fat-infiltrated patients.

Proposed method

• The method introduces an asymmetric cycle-consistency loss that modifies the standard cycle-consistency objective to handle cases where multiple images in the source domain map to a single image in the target domain.
• Instead of enforcing symmetry in reconstruction, the loss prioritizes the direction where the mapping is many-to-one, reducing ambiguity in the translation process.
• The loss function is derived by reweighting the cycle-consistency term based on the likelihood of mapping ambiguity, favoring more reliable reconstruction paths.
• The approach is integrated into a GAN-based image translation framework without altering the network architecture or increasing computational overhead.
• The method is trained end-to-end on unpaired thigh MRI scans, with the goal of translating fat-infiltrated muscle images to a reference healthy state.

Experimental results

Research questions

• RQ1 How does asymmetric cycle-consistency loss improve image translation performance in the presence of many-to-one mappings in medical imaging?
• RQ2 Can the proposed loss enhance segmentation accuracy of muscles in fat-infiltrated MRI scans without architectural changes?
• RQ3 How does the asymmetric loss compare to standard symmetric cycle-consistency in handling ambiguous mappings from pathological to healthy images?
• RQ4 Does the method maintain training stability and convergence speed compared to baseline GANs with symmetric cycle-consistency?

Key findings

• The proposed asymmetric cycle-consistency loss significantly improves image translation quality on thigh MRI scans with fat-infiltrated muscles.
• The method achieves better segmentation performance of muscle regions in fat-infiltrated patients compared to standard cycle-consistency training.
• The improvement is achieved without modifying the network architecture or increasing computational complexity.
• Empirical results show reduced reconstruction error and more plausible translations in ambiguous mapping scenarios.