Hypothesis transfer in medical image analysis

Recent years have seen a wealth of clinical evidence accumulate in favor of the radiomics hypothesis, wherein standard-issue oncological medical images, such as CTs, PET-CTs and MRIs, exhibit reliable fingerprints of cancer tumor genetic makeup, enabling predictions of patient prognosis, treatment resistance and side effects to be made directly from an analysis of imaging data. However, these techniques require a large number of detailed hand-labeled data and annotations and so far have relied on hand-engineered features, limiting their scope and practical impact. This project aims at leveraging labels obtained from different datasets using novel domain adaptation techniques.
In vision applications, Hypothesis Transfer Learning (HTL) has emerged as a learning strategy to adapt the knowledge learned from one dataset (source domain) to a query dataset (target domain) and has shown superior performance to more conventional domain adaptation methods [1]. HTL has only been experimented with covariate shift (domain shift) and assumes similar class distributions between source and target domains. However, class distribution shifts and feature distribution shifts between datasets are common phenomena in medical imaging. With this incentive, an experimental study on the effectiveness of HTL in the medical domain and how it contrasts with other domain adaptation applications will be conducted.

Faculty Supervisor:

Thomas Fevens

Student:

Partner:

Imagia

Discipline:

Computer science

Sector:

Information and cultural industries; Professional, scientific and technical services

University:

Concordia University

Program: