Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Transfer learning techniques for application of deep neural networks in 3D fluorescent microscopy images
Application of deep neural networks to 3D fluorescent microscopy datasets poses some specific challenges such as differences in their input distributions or scarce labeled data. We aim to address them by investigating transfer learning techniques such as domain adaptation of multi-task learning.
Keywords: transfer learning; neural networks; domain adaptation; multi-task learning; deep learning; image segmentation; 3D microscopy; fluorescent microscopy
3D fluorescent microscopy is becoming a standard image modality in most biomedical research labs, as it allows to produce very detailed images of various cellular populations on biological tissues. With vast amounts of data already available, the challenge resides in the generation of tools for automatic image segmentation and analysis.
Deep neural networks can address these problems when enough labeled images exist and their intensities follow a similar distribution. However, fluorescent microscopy images usually have a different input probability distribution, as they are very susceptible to the sample preparation and image acquisition protocols. Furthermore, labeled data is scarce, as interpretation of these images requires a lot of experience. To date, no methods exist that successfully generalize deep learning techniques in this type of images.
3D fluorescent microscopy is becoming a standard image modality in most biomedical research labs, as it allows to produce very detailed images of various cellular populations on biological tissues. With vast amounts of data already available, the challenge resides in the generation of tools for automatic image segmentation and analysis.
Deep neural networks can address these problems when enough labeled images exist and their intensities follow a similar distribution. However, fluorescent microscopy images usually have a different input probability distribution, as they are very susceptible to the sample preparation and image acquisition protocols. Furthermore, labeled data is scarce, as interpretation of these images requires a lot of experience. To date, no methods exist that successfully generalize deep learning techniques in this type of images.
In this project, transfer learning methods will be explored to overcome the image segmentation challenges posed by fluorescent microscopy datasets. We will explore how the performance of the neural networks is affected by the input distribution of the images and how to overcome the limitations of scarce labeled data. For this purpose, we will investigate methods in the field of domain adaptation and multi-task learning.
In this project, transfer learning methods will be explored to overcome the image segmentation challenges posed by fluorescent microscopy datasets. We will explore how the performance of the neural networks is affected by the input distribution of the images and how to overcome the limitations of scarce labeled data. For this purpose, we will investigate methods in the field of domain adaptation and multi-task learning.