16 January 2025 in IEEE Transactions on Medical Imaging
Cataract surgery, a widely performed operation worldwide, is incorporating semantic segmentation to advance computer-assisted intervention. However, the tissue appearance and illumination in cataract surgery often differ among clinical centers, intensifying the issue of domain shifts. While domain adaptation offers remedies to the shifts, the necessity for data centralization raises additional privacy concerns. To overcome these challenges, we propose a Multi-view Test-time Adaptation algorithm (MUTA) to segment cataract surgical scenes, which leverages multi-view learning to enhance model training within the source domain and model adaptation within the target domain. In the training phase, the segmentation model is equipped with multi-view decoders to boost its robustness against variations in cataract surgery. During the inference phase, test-time adaptation is implemented using multi-view knowledge distillation, enabling model updates in clinics without data centralization or privacy concerns. We conducted experiments in a simulated cross-center scenario using several cataract surgery datasets to evaluate the effectiveness of MUTA. Through comparisons and investigations, we have validated that MUTA effectively learns a robust source model and adapts the model to target data during the practical inference phase. Code and datasets are available at https://github.com/liamheng/CAI-algorithms.
14 July 2025 in Physics in Medicine & Biology
Clinical medical images often suffer from compromised quality, which negatively impacts the diagnostic process by both clinicians and AI algorithms. While GAN-based enhancement methods have been commonly developed in recent years, delicate model training is necessary due to issues with artifacts, mode collapse, and instability. Diffusion models have shown promise in generating high-quality images superior to GANs, but challenges in training data collection and domain gaps hinder applying them for medical image enhancement. Additionally, preserving fine structures in enhancing medical images with diffusion models is still an area that requires further exploration. To overcome these challenges, we propose generalizable medical image enhancement using structure-preserved diffusion models (GEDM) leverages joint supervision from enhancement and segmentation to boost structure preservation and generalizability. Specifically, synthetic data is used to collect high-low quality paired training data with structure masks, and the Laplace transform is employed to reduce domain gaps and introduce multi-scale conditions. GEDM conducts medical image enhancement and segmentation jointly, supervised by high-quality references and structure masks from the training data. Four datasets of two medical imaging modalities were collected to implement the experiments, where GEDM outperformed state-of-the-art methods in image enhancement, as well as follow-up medical analysis tasks.
1 January 2025 in Computational Biomedicine
Aims: Medical image translation is widely used for data augmentation and cross-domain adaptation in clinical image analysis. However, the nature of medical imaging makes it challenging to collect high-quality samples for the training of translation models. Because of the limited access to and costly expense of medical images, distribution bias is commonly observed between the source and target samples, and this finally leads the models to mismatch the target domain. Methods: To promote medical image translation, a bias correction method, named content adaptation, has been proposed in this study to align the training samples in the data space. Based on the invariant medical topological structure, paired samples are constructed from weakly paired and unpaired data using content adaptation to correct the distribution bias and promote the image-to-image translation. Results: Experiments on retinal fundus image translation and COVID-19 CT synthesis demonstrate that the proposed method effectively suppresses structural hallucination and improves both visual quality and quantitative performance. Consistent gains are observed across multiple backbone models under different supervision settings. The results suggest that explicit anatomical alignment provides an effective and model-agnostic way to mitigate distribution bias in medical image translation. By bridging weakly paired data with paired translation paradigms, the proposed approach enhances structural fidelity without requiring strong supervision. Conclusion: This work presents a topology-guided content adaptation strategy that improves robustness and reduces hallucination in medical image translation. The proposed framework is general and can be readily integrated into existing translation models, offering a practical solution for data-scarce medical imaging scenarios.