| KEYWORD |
Generative techniques for high-resolution medical image synthesis
keywords COMPUTER VISION, DEEP LEARNING, GENERATIVE ADVERSARIAL NETWORKS, MEDICAL IMAGING
Reference persons LIA MORRA
Research Groups DAUIN - GR-09 - GRAphics and INtelligent Systems - GRAINS
Description This thesis falls within the activities of the AIBIBANK (Bio-Banking for Artificial Intelligence) project, carried out in collaboration with HealthTriage srl (I3P). The overarching goal is to develop artificial intelligence solution for the automated triage of digital mammography images in order to focus the radiologists’ attention to potentially normal cases.
Since cancer is a rare disease, the dataset available is highly skewed towards normal cases. In addition, the content of mammography images depend on the vendor (manufacturer) of the mammographic unit, and the characteristics of the woman (age, breast density, etc.). In order to make the models more robust to under-represented categories, the aim of this thesis is the development of generative models capable of synthesizing high resolution images. One or more of the following objectives will be pursued:
- developing models able to convert (with techniques such as CycleGAN) the global characteristics of the image, such as changing the vendor, the breast side or density
- generate and inject synthetic lesions in normal cases, from the same or different vendors, using inpanting techniques. Multiple views of the same lesion should be generated.
- experiment with the use of models based on Stable Diffusion for the generation of high resolution mammography images
Strong programming and analytical skills are required. Experience with deep learning and Pytorch is required.
Suggested readings
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6430964/pdf/JMI-006-031411.pdf
https://www.frontiersin.org/articles/10.3389/fonc.2022.868257/full https://arxiv.org/pdf/2209.09809.pdf
Required skills machine learning, deep learning, Pytorch
Deadline 20/02/2024
PROPONI LA TUA CANDIDATURA