Generating semantic-aware digital health images using generative models

AI Quick Summary

Generating semantic-aware and plausible digital health images by using generative models combined with a-priori knowledge Initial situation and problem definition:While the literature has shown promising efforts towards the automatic analysis of digital health images, such as organ segmentation or tissue detection, existing approaches rely on the availability of patient data together with its annotation. Due to privacy issues, real patient data is usually scarce and its annotation demains expert knowledge and represents a time-consuming task. To cope with these issues, this project intends the study and development of generative approaches for the generating of plausible and real-looking patient images.Aim of the work and expected results:The aim of this work is to review existing generative approaches for image generation and their applicability applicability into the digital health domain. The main goal is to develop a generative model that is capable of generating realistic-looking and semantic plausible digital health images. This work has three main outcomes. First, we focus on studying existing generative approaches for image generation and on reproducing their results. Second, we target the development of novel modules/blocks by adding domain-specific knowledge.Desired methods and procedure:The first stage of the project involves the selection of the proper medical datasets. Next, we focus on reproducing existing approaches from available github repositories. The goal is to understand the limitations of existing methods. The second stage will involve tackling the weaknesses of existing approaches by exploring different blocks/modules and also by using knowledge from the medical literature. Every week, there will be a meeting, where the student will provide insights of the achievements via presentations, jupyter notebooks, etc. Discussions about next steps and guidance will be provided by the mentor. The source code will be developed in python/pytorch and will be privately stored in github. The. code can be built from scratch or from existing repositories. The bachelor work will be written in parallel as the experiments are being conducted.Creativity, variants, innovation: As most of the existing generative models were evaluated in general object recognition or segmentation tasks, there is room for studying and adapting existing approaches in the context of digital health. Thus, this topic brings different exploration possibilities, while still being challenging. Besides the technical aspects, it gives the chance to the student to learn more about the digital health domain.

Project Idea Metadata

• Project Idea Name: Generating semantic-aware digital health images using generative models
• Date: 8/25/2022 4:33:00 AM
• Administrators:

Project Idea Description

Initial situation and problem definition:

While the literature has shown promising efforts towards the automatic analysis of digital health images, such as organ segmentation or tissue detection, existing approaches rely on the availability of patient data together with its annotation. Due to privacy issues, real patient data is usually scarce and its annotation demains expert knowledge and represents a time-consuming task. To cope with these issues, this project intends the study and development of generative approaches for the generating of plausible and real-looking patient images.

Aim of the work and expected results:

The aim of this work is to review existing generative approaches for image generation and their applicability applicability into the digital health domain. The main goal is to develop a generative model that is capable of generating realistic-looking and semantic plausible digital health images. This work has three main outcomes. First, we focus on studying existing generative approaches for image generation and on reproducing their results. Second, we target the development of novel modules/blocks by adding domain-specific knowledge.

Desired methods and procedure:

The first stage of the project involves the selection of the proper medical datasets. Next, we focus on reproducing existing approaches from available github repositories. The goal is to understand the limitations of existing methods. The second stage will involve tackling the weaknesses of existing approaches by exploring different blocks/modules and also by using knowledge from the medical literature. Every week, there will be a meeting, where the student will provide insights of the achievements via presentations, jupyter notebooks, etc. Discussions about next steps and guidance will be provided by the mentor. The source code will be developed in python/pytorch and will be privately stored in github. The. code can be built from scratch or from existing repositories. The bachelor work will be written in parallel as the experiments are being conducted.

Creativity, variants, innovation:

As most of the existing generative models were evaluated in general object recognition or segmentation tasks, there is room for studying and adapting existing approaches in the context of digital health. Thus, this topic brings different exploration possibilities, while still being challenging. Besides the technical aspects, it gives the chance to the student to learn more about the digital health domain.

Generating semantic-aware and plausible digital health images by using generative models combined with a-priori knowledge