Generation and evaluation of synthetic image sequences for prediction in the biomedical field

Resum

This thesis focuses on the application of image and video generation models to predict the evolution of a biological element in biomedical contexts. The goal is to generate a sequence of images that represent the different stages through which a case study will evolve. Through a comprehensive analysis of these models, their potential to generate coherent biological developmental sequences will be explored. During the thesis work, the lack of metrics to evaluate the coherence of the sequences was identified, leading to the development of new statistical measures based on the analysis of stages in the generated videos. Using these metrics, a new model is created that demonstrates superior stage coherence in sequence generation. Giving the limitations of GAN-based models in effectively using real input images as conditions, the study shifts to architectures based on diffusion. This shift facilitates the performance of image sequence prediction tasks. The incorporation of the proposed metrics allows the generation of different predictions depending on the number of real images used for their generation, taking into account the stage coherence and the detection of movements or abrupt changes in neighboring images. Finally, this work also proposes a new model based on a siamese architecture to extend the duration of videos without degrading image quality. This leads to a significant increase in video length, thereby improving the detail of its evolution. Regarding future work, the primary goal is to improve image quality by increasing the resolution and the level of detail generated. In addition, the inclusion of new statistical measures and data from the studied cases will be explored to further benefit various biomedical analyses. This extension could pave the way for the analysis of tumor evolution, skin lesion growth, and other cases that require in-depth study of their evolution.