Estimating Medical Indicators for Disease Diagnosis and Tracking with Deep Regression Models

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Estimating Medical Indicators for Disease Diagnosis and Tracking with
Deep Regression Models"

By

Mr. Weihang DAI


Abstract:

Deep regression problems involve estimating continuous, real-number variables
from unstructured inputs such as images or videos. These problems are important
for medical imaging analysis since disease tracking and diagnosis are often
based on medical indicators that take on real-number values. Despite their
practical importance, deep regression is not as well explored as classification
and segmentation tasks In this thesis, we propose state-of-the-art methods for
deep regression in medical imaging analysis by addressing their common
characteristics and challenges. Specifically, our methods are based on three
general approaches: improving feature representations for deep regression,
using unlabeled data through semi-supervision, and using region-of-interest
segmentations for additional context.

We first propose a novel adaptive contrastive learning framework, AdaCon, for
improved feature learning. Existing contrastive learning methods for
classification cannot be directly applied to regression as they cannot account
for label distance between samples. AdaCon allows features to reflect distance
relationships however, which improves downstream regression performance.

We then examine semi-supervised methods to address the challenge of limited
annotations for medical data. We propose Uncertainty-Consistent Variational
Model Ensembling (UCVME), which uses uncertainty estimates for unlabeled data
to focus training on high-quality regression pseudo-labels. We also propose
Contrastive Learning with Spectral Seriation (CLSS), which extends contrastive
learning for regression to a semi-supervised setting.

Finally, we explore how region-of-interest segmentations can provide additional
context for regression. We propose cyclical self-supervision (CSS), which
generates improved left-ventricle segmentation predictions for input into
ejection fraction regression models. We also show how segmentation masks can be
used to extract radiomic features, which can then be combined with deep
features using our radiomics informed deep learning (RIDL) framework.

We demonstrate our methods on a variety of medical tasks and outperform
existing approaches. Our methods have direct clinical benefits as it allows for
more reliable indicators readings to be obtained for improved diagnosis.


Date:                   Thursday, 27 July 2023

Time:                   8:45am - 10:45am

Venue:                  Room 3494
                        lifts 25/26

Chairperson:            Prof. Mike SO (ISOM)

Committee Members:      Prof. Tim CHENG (Supervisor)
                        Prof. Xiaomeng LI (Supervisor)
                        Prof. Hao CHEN
                        Prof. Long QUAN
                        Prof. Yanglong LU (MAE)
                        Prof. Hongsheng LI (CUHK)


**** ALL are Welcome ****