Incorporating privileged information in the context of deep learning for medical imaging

In this PhD project we aim to design an algorithm that is specifically designed for incorporating expert medical knowledge (privileged information) into the learning course of image classification deep learning models. It is worth mentioning that the privileged information will only be available during the training phase along with the classification label of each image but will be absent during the testing phase. A similar problem has been investigated in [5] where the privileged information was given in the form of a segmentation mask and was incorporated into the training stage to improve the performance of Convolutional Neural Networks (CNNs). Another approach in [6] used a heteroscedastic dropout approach to incorporate images as privileged information in the learning phase of CNNs and Recurrent Neural Networks (RNNs). Unlike the exiting approaches which are limited to integrating imaging data as privileged information, in this project we aim to incorporate various types and formats of privileged information (such as texts, images and numerical features), depending on the knowledge domain, to improve the performance of image classification deep learning models. The framework that is to be developed in this project will be applied and tested in the context of medical imaging practical problems.

More information on the project, from potential impact to references, can be found on the accompanying PDF.

To apply, please complete the project proposal form and the online application.

Only local Hubs members can access this page. Join the community today: https://archive.phdhub.eu/register