Deep Regression for Medical Imaging Analysis with Improved Feature Learning and Label-Efficient Training Techniques

PhD Thesis Proposal Defence


Title: "Deep Regression for Medical Imaging Analysis with Improved Feature
Learning and Label-Efficient Training Techniques"

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
particularly important for medical imaging analysis since disease tracking and
diagnosis are often based on medical indicators which take on continuous
real-number values, such as bone mineral density for osteoporosis or ejection
fraction for cardiac disease. Unlike classification and segmentation problems
however, deep regression problems are not as well explored despite their
practical applications. This thesis addresses common challenges associated with
deep regression problems, specifically ways to improve feature learning,
annotation-efficient learning methods, and usage of additional context to
improve regression predictions. By addressing these three areas of improvement,
we are able to achieve better performance on different medical imaging analysis
tasks, which can lead to more accurate downstream diagnosis and assessment for
patient treatment.

First, we address the problem of feature learning for regression tasks by
proposing AdaCon, an adaptive contrastive learning framework that can be
applied to deep regression problems. Prior works such as SimCLR, MoCo, and
SupCon have shown that contrastive learning methods can improve feature
learning and surpass state-of-the-art performance on classification tasks.
These methods cannot be directly applied to regression problems however, as
regression labels reflect label distance information unlike individual class
labels, and existing contrastive loss functions do not account for this. We
propose the first contrastive learning framework for deep image regression,
namely AdaCon, which incorporates label distance relationships as part of the
learned feature representations. Our method allows for better performance in
downstream regression tasks and can be used as a plug-and-play module to
improve performance of existing regression methods.

Second, we propose UCVME, a semi-supervised learning approach to deep
regression. Semi-supervised learning for classification problems are well
explored, but less attention has been given to deep regression problems.
Furthermore, methods such as confidence thresholding, which can be applied on
probability outputs for classification problems, cannot be directly used for
regression. Our work proposes a novel approach to semi-supervised regression,
namely Uncertainty-Consistent Variational Model Ensembling (UCVME), which
improves training by generating high-quality pseudo-labels and uncertainty
estimates for heteroscedastic regression. Our method is based on the
observation that aleatoric uncertainty is only dependent on input data by
definition and should be equal for the same inputs. By enforcing consistency on
uncertainty estimates from co-trained models, we can significantly improve the
quality of uncertainty estimates, thus allowing higher quality pseudo-labels to
be assigned greater importance under heteroscedastic regression.

Third, we propose cyclical self-supervision (CSS), a novel unsupervised loss
that can be used to enforce cyclicality for sequential data. We use the
cyclical nature of the heartbeat as a prior and enforce CSS as an additional
constraint to perform semi-supervised left-ventricle segmentation from
echocardiogram video sequences. The segmentation predictions are then used as
additional input to provide context for LVEF regression, which allows knowledge
learned from predicting LV segmentations to be incorporated into LVEF
regression. Furthermore, we introduce teacher-student distillation to distill
the information from LV segmentation masks into an end-to-end LVEF regression
model such that only video inputs are required.

Finally, we introduce radiomics-informed deep learning (RIDL), a framework that
combines context from segmentation mask inputs, deep features, and radiomic
features for downstream tasks such as disease classification and prognosis.
Common radiomic shape features, such as cancer tumor size, are sometimes used
directly as medical indicators for diagnosis and tracking. Our work shows how
these different sources of information can be combined to perform other
downstream medical tasks through a novel approach that combines the advantages
of deep learning and radiomic modelling. Unlike existing hybrid techniques that
mostly rely on naive feature concatenation, we observe that radiomic feature
selection methods can serve as an information prior, and propose supplementing
low-level deep neural network (DNN) features with locally computed radiomic
features. This reduces DNN over-fitting and allows local variations between
radiomic features to be better captured. Furthermore, we ensure complementary
information is learned by deep and radiomic features by designing a novel
feature de-correlation loss.


Date:                   Monday, 15 May 2023

Time:                   3:00pm - 5:00pm

Venue:                  Room 5510
                        lifts 25/26

Committee Members:      Prof. Tim Cheng (Supervisor)
                        Dr. Xiaomeng Li (Supervisor)
                        Prof. Pedro Sander (Chairperson)
                        Dr. Hao Chen


**** ALL are Welcome ****