A User-Interface for Automatic Generation of Digital Twins in Radiology : Exploring Selection Mechanisms in Curriculum-Learning for training Generative Adversarial Networks (GAN) in CT-to-PET Image Translation

Sammanfattning

Despite the tremendous advances that deep learning has contributed to the synthetic medical imaging domain, availability and practicality limit the extent to which practitioners can test, evaluate, and use state-of-the-art generative models. Although calculated metrics generally exhibit an adequate measure when evaluating model performance, visual inspection remains crucial to evaluating the real-world utility of synthetic medical images. Training generative models is resource intensive and unstable. Curriculum-Learning offers a way to increase accuracy by controlling the order and frequency of data exposure, but the choice of data selection strategy impacts model performance and computational cost. To address both the accessibility of generative model output and the efficiency of training, this work presents: (i) a User-Interface to explore machine learning (ML) methods to generated synthetic medical images, and (ii) an empirical study on the impact of data selection mechanisms in a curriculum-driven training paradigm. The User-Interface was developed following a user-centered, iterative design paradigm to facilitate gradual convergence towards user needs. Meanwhile, the Curriculum-Learning framework was extended with different dynamic sampling strategies that alter the training data distribution over time. Results show a tradeoff between computational time and performance across strategies, emphasizing the role of qualitative data and model generalization. Together, these contributions provide an accessible and transparent tool for exploring ML-generated medical images and offer insights into optimizing training workflows. By integrating generative image-to-image translation models with an interface utilizing user-centered design, we aim to enhance diagnostic precision, and reduce the need for repeated imaging, and minimize patient radiation exposure.