ºÚÁϸ£ÀûÍø

In her dissertation, Evi Huijben builds on existing insights into generative AI and investigates how this technology can be applied to address key challenges in medical image analysis. Through a series of experimental models, she demonstrates that generative AI is not only capable of producing realistic medical images, but also of enhancing and supplementing existing datasets.

Main Findings

• Conditional image synthesis offers a way to generate specific images required for downstream tasks in medical settings, where acquiring the desired data is often expensive, impractical, or even impossible.
• Structuring latent spaces within AI models enables targeted control over the generated outputs, which is important for medical image synthesis to ensure the images are clinically relevant. 
• Labeled data are essential for training and evaluating supervised or conditional deep learning models. However, labeled data are often scarce in medical imaging, and thus, the development of pipelines that can function with a small amount of labeled or weakly labeled data is crucial.

Applications and Methodological Insights

Huijben's research introduces several AI methods for generating medical images across various applications, image types, and anatomical regions.

She used generative AI to create synthetic datasets to train classification models. One model disentangles class characteristics and generates images tailored to particular patient groups, while another synthesizes images of rare conditions that are underrepresented in existing datasets, thereby improving their detection.

In addition, generative AI was used to produce patient-specific synthetic images. One model predicts disease progression over time – for example, by generating a future brain MRI to support prognosis in Alzheimer's disease – while another detects abnormalities in brain MRIs, such as tumors or artifacts caused by patient movement during scanning. Finally, Huijben organized the international competition SynthRAD2023, which focused on generating synthetic CT images from other imaging types, such as MRIs, enabling a fair comparison of models through a large dataset of cancer patients.

Challenges

Despite its potential, generative AI in medical images faces several challenges.

• Model design is complex and often constrained by available computational resources, meaning that the models are not always optimized for clinical relevance.
• Evaluating synthetic images is difficult due to limited ground truth images and the lack of clinically interpretable evaluation metrics. While humans can easily identify visual artifacts, creating objective and reliable quality metrics for medical images remains an open problem.
• Facilitating clinical translation is essential for advancing generative AI models toward real-world applications. Achieving this, however, requires careful metric selection, expert validation, clear acceptance criteria, and diverse benchmark datasets.

Conclusion

This dissertation highlights that, with task-specific design and careful validation, generative models can create realistic images and hold potential for clinical use. This work contributes to the growing integration of AI in healthcare and could ultimately improve patient outcomes.