The main objective of this thesis is to explore and compare methods of image generation of cervical cytological images. This dissertation will explore multi-object generation while focusing on the exploration of the metrics and strategies which guarantee that the synthetic images are not just visually appealing but contain meaningful features and maintain the reliability of images. This dissertation will rely on public datasets, as well as private dataset collected in the scope of CLARE and TAMI projects.
In particular, the explored data augmentation techniques will contribute to the improvements of the currently available cervical image datasets by: i) increasing the overall volume of training data; ii) extending the number of samples of the underrepresented classes (especially squamous cell carcinoma); and iii) balancing the volume disparities between the data classes.
The impact of the explored data augmentation techniques will be evaluated in two different ways: i) improvement of already developed DSS's by training with the augmented dataset (the student will have access to already trained AI models for CC screening, developed in the ambit of the TAMI project); and ii) acceptance of the AIgenerated images by medical experts in terms of realism and relevance for the respective classes.
Expected direct impact in task execution of current and future Computer Vision projects, such as TAMI. Additionally, it will also contribute to the dissemination indicators of the referred funded project.
Author: Tiago Alves
Type: MSc thesis
Partner: FEUP – Faculdade de Engenharia da Universidade do Porto