Abstract:The presence of domain shift in medical imaging is a common issue, which can greatly impact the performance of segmentation models when dealing with unseen image domains. Adversarial-based deep learning models, such as Cycle-GAN, have become a common model for approaching unsupervised domain adaptation of medical images. These models however, have no ability to enforce the preservation of structures of interest when translating medical scans, which can lead to potentially poor results for unsupervised domain adaptation within the context of segmentation. This work introduces the Structure Preserving Cycle-GAN (SP Cycle-GAN), which promotes medical structure preservation during image translation through the enforcement of a segmentation loss term in the overall Cycle-GAN training process. We demonstrate the structure preserving capability of the SP Cycle-GAN both visually and through comparison of Dice score segmentation performance for the unsupervised domain adaptation models. The SP Cycle-GAN is able to outperform baseline approaches and standard Cycle-GAN domain adaptation for binary blood vessel segmentation in the STARE and DRIVE datasets, and multi-class Left Ventricle and Myocardium segmentation in the multi-modal MM-WHS dataset. SP Cycle-GAN achieved a state of the art Myocardium segmentation Dice score (DSC) of 0.7435 for the MR to CT MM-WHS domain adaptation problem, and excelled in nearly all categories for the MM-WHS dataset. SP Cycle-GAN also demonstrated a strong ability to preserve blood vessel structure in the DRIVE to STARE domain adaptation problem, achieving a 4% DSC increase over a default Cycle-GAN implementation.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the domain shift problem in the field of medical images, which is a common challenge in medical image segmentation tasks. Specifically: 1. **Impact of Domain Shift**: When dealing with unseen image domains, domain shift can significantly affect the performance of segmentation models. For example, brain MRI images generated by different scanners may have completely different intensity distributions, resulting in poor model performance when these images are used jointly. 2. **Limitations of Existing Methods**: Although deep - learning models based on adversarial learning (such as Cycle - GAN) have been widely used for unsupervised domain adaptation (UDA), these models cannot effectively maintain the integrity of the structures of interest during the image conversion process, which may lead to poor results. ### Solution To solve the above problems, this paper proposes **Structure Preserving Cycle - GAN (SP Cycle - GAN)**, and its main improvements are as follows: - **Introducing Segmentation Loss Term**: By adding a segmentation loss term during the Cycle - GAN training process, the integrity of important structures is ensured during image translation. Specifically, SP Cycle - GAN uses the U - Net model for simultaneous training and strengthens the protection of structures through the Focal - Tversky loss function. - **Experimental Verification**: The paper conducts experimental verification through multiple datasets, including retinal vessel segmentation tasks on the STARE and DRIVE datasets, and multi - class segmentation tasks of the left ventricle (LV) and myocardium on the MM - WHS dataset. The experimental results show that SP Cycle - GAN outperforms traditional Cycle - GAN and other baseline methods in multiple evaluation metrics. ### Formula Representation To ensure structure preservation, SP Cycle - GAN introduces the following formulas: - **Tversky Index**: \[ TI_c=\frac{\sum_{i = 1}^{N}p_i^c g_i^c+\epsilon}{\sum_{i = 1}^{N}p_i^c g_i^c+\alpha\sum_{i = 1}^{N}p_i^{\bar{c}}g_i^c+\beta\sum_{i = 1}^{N}p_i^c g_i^{\bar{c}}+\epsilon} \] where \(p_i^c\) and \(g_i^c\) represent the probabilities that the \(i\)-th pixel in the prediction and the ground - truth label belongs to the foreground class respectively, and \(\alpha\) and \(\beta\) are parameters used for flexibly punishing false negatives and false positives. - **Focal - Tversky Loss**: \[ FTL_c=\sum_c(1 - TI_c)^{\gamma} \] where \(\gamma\) is the focus term, which is used to focus on low - confidence samples. - **Structure - Preserving Loss**: \[ L_{Struct}(U,G,F,X,L)=FTL_c(L,U(F(G(X)))) \] - **Total Loss Function**: \[ L_{SGAN}(G,F,D_X,D_Y)=L_{GAN}(G,D_Y,X,Y)+L_{GAN}(F,D_X,X,Y)+L_{cyc}(G,F)+\zeta\cdot L_{Struct}(U,G,F,X,L) \] Through these improvements, SP Cycle - GAN can better preserve the key structures in medical images in unsupervised domain adaptation tasks, thereby improving the segmentation performance.

Structure Preserving Cycle-GAN for Unsupervised Medical Image Domain Adaptation

SA-GAN: Structure-Aware GAN for Organ-Preserving Synthetic CT Generation

Multilevel structure-preserved GAN for domain adaptation in intravascular ultrasound analysis

Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks

Unsupervised Domain Adaptation via CycleGAN for White Matter Hyperintensity Segmentation in Multicenter MR Images

Edge-preserving Domain Adaptation for semantic segmentation of Medical Images

SUSAN: Segment Unannotated image Structure using Adversarial Network

Unpaired Brain MR-to-CT Synthesis using a Structure-Constrained CycleGAN

Unsupervised MR-to-CT Synthesis Using Structure-Constrained CycleGAN.

Fundus Image Enhancement Via Semi-Supervised GAN and Anatomical Structure Preservation

Structure-preserving image translation for multi-source medical image domain adaptation

Structure and Illumination Constrained GAN for Medical Image Enhancement

CycleSGAN: A cycle-consistent and semantics-preserving generative adversarial network for unpaired MR-to-CT image synthesis

Dual-Channel in Spatial-Frequency Domain CycleGAN for perceptual enhancement of transcranial cortical vascular structure and function

Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks

Anatomy-constrained synthesis for spleen segmentation improvement in unpaired mouse micro-CT scans with 3D CycleGAN

Generalisable Cardiac Structure Segmentation via Attentional and Stacked Image Adaptation

Structure-Preserving Synthesis: MaskGAN for Unpaired MR-CT Translation

Shape-Consistent Generative Adversarial Networks for Multi-Modal Medical Segmentation Maps

Shape-aware synthesis of pathological lung CT scans using CycleGAN for enhanced semi-supervised lung segmentation

StAC-DA: Structure aware cross-modality domain adaptation framework with image and feature-level adaptation for medical image segmentation