PhD topic proposal: MAP-i
Scientific Area:
Computer Vision, Machine/Deep Learning, Segmentation, Transfer Learning, Low SNR, blood vessels
Description:
The segmentation and analysis of blood vessels in medical images have been allowing clinicians to assess several health conditions that directly impact these structures (such as aneurysms and stenosis) and also to obtain clues on some systemic diseases (for example, diabetes and hypertension).
Unfortunately, the study of vascular networks is typically a time consuming and repetitive procedure.
Computer vision has been playing an important role in supporting the segmentation and analysis of blood vessels, increasing the efficiency of medical care. Despite the huge progress that has been achieved since the first works, blood vessel segmentation is still far from matching the rationale used by human experts and is still lacking in some aspects.
A large portion of the related literature comprises algorithms focusing a given blood vessel segmentation scenario (a combination of a particular vascular tree and imaging acquisition protocol).
Targeting a specific use case allows to more heavily rely on prior knowledge and design more optimized methodologies.
Two application-specific algorithms could be proposed. The first one concerns a framework tailored for the extraction and analysis of the Deep Inferior Epigastric Perforators, blood vessels lying in the anterior region of the abdominal wall. Their identification and characterization are essential for the preoperative planning of the state-of-the-art flaps using tissue from the belly to reconstruct the breast. In order to guarantee that the new breast is adequately vascularized after re-anastomosis, the surgeons need to include blood vessels with proper characteristics.
The second one targets the segmentation of blood vessels in retinal images, where competitive performance is achieved using a much faster deep neural network design, which is relevant for screening programs, where a lot of data are generated in a short amount of time.
The collaboration with Champalimaud Foundation in previous projects allowed the collection of a specific dataset and the correspondent annotation of the data that can be used in the current proposal.
References:
https://www.sciencedirect.com/science/article/pii/S0895611118305810?via%3Dihub https://www.sciencedirect.com/science/article/pii/S0960977620300023
https://dl.acm.org/doi/10.1007/978-3-030-32239-7_11
Supervisor:
Hélder Filipe Pinto de Oliveira, University of Porto, INESC TEC
Co-Supervisor:
Tania Pereira, INESC TEC
Host Institution:
INESC TEC
External reviewer:
Luís Alexandre, UBI