Symbiotic deep learning for medical image analysis with applications in real-time diagnosis for fetal ultrasound screening
File(s)
Author(s)
Budd, Samuel
Type
Thesis or dissertation
Abstract
The last hundred years have seen a monumental rise in the power and capability of machines to
perform intelligent tasks in the stead of previously human operators. This rise is not expected
to slow down any time soon and what this means for society and humanity as a whole remains
to be seen. The overwhelming notion is that with the right goals in mind, the growing influence
of machines on our every day tasks will enable humanity to give more attention to the truly
groundbreaking challenges that we all face together. This will usher in a new age of human
machine collaboration in which humans and machines may work side by side to achieve greater
heights for all of humanity. Intelligent systems are useful in isolation, but the true benefits of
intelligent systems come to the fore in complex systems where the interaction between humans
and machines can be made seamless, and it is this goal of symbiosis between human and machine
that may democratise complex knowledge, which motivates this thesis. In the recent past, datadriven
methods have come to the fore and now represent the state-of-the-art in many different
fields. Alongside the shift from rule-based towards data-driven methods we have also seen a
shift in how humans interact with these technologies. Human computer interaction is changing
in response to data-driven methods and new techniques must be developed to enable the same
symbiosis between man and machine for data-driven methods as for previous formula-driven
technology.
We address five key challenges which need to be overcome for data-driven human-in-the-loop
computing to reach maturity. These are (1) the ’Categorisation Challenge’ where we examine
existing work and form a taxonomy of the different methods being utilised for data-driven
human-in-the-loop computing; (2) the ’Confidence Challenge’, where data-driven methods must
communicate interpretable beliefs in how confident their predictions are; (3) the ’Complexity
Challenge’ where the aim of reasoned communication becomes increasingly important as the
complexity of tasks and methods to solve also increases; (4) the ’Classification Challenge’ in
which we look at how complex methods can be separated in order to provide greater reasoning
in complex classification tasks; and finally (5) the ’Curation Challenge’ where we challenge the
assumptions around bottleneck creation for the development of supervised learning methods.
perform intelligent tasks in the stead of previously human operators. This rise is not expected
to slow down any time soon and what this means for society and humanity as a whole remains
to be seen. The overwhelming notion is that with the right goals in mind, the growing influence
of machines on our every day tasks will enable humanity to give more attention to the truly
groundbreaking challenges that we all face together. This will usher in a new age of human
machine collaboration in which humans and machines may work side by side to achieve greater
heights for all of humanity. Intelligent systems are useful in isolation, but the true benefits of
intelligent systems come to the fore in complex systems where the interaction between humans
and machines can be made seamless, and it is this goal of symbiosis between human and machine
that may democratise complex knowledge, which motivates this thesis. In the recent past, datadriven
methods have come to the fore and now represent the state-of-the-art in many different
fields. Alongside the shift from rule-based towards data-driven methods we have also seen a
shift in how humans interact with these technologies. Human computer interaction is changing
in response to data-driven methods and new techniques must be developed to enable the same
symbiosis between man and machine for data-driven methods as for previous formula-driven
technology.
We address five key challenges which need to be overcome for data-driven human-in-the-loop
computing to reach maturity. These are (1) the ’Categorisation Challenge’ where we examine
existing work and form a taxonomy of the different methods being utilised for data-driven
human-in-the-loop computing; (2) the ’Confidence Challenge’, where data-driven methods must
communicate interpretable beliefs in how confident their predictions are; (3) the ’Complexity
Challenge’ where the aim of reasoned communication becomes increasingly important as the
complexity of tasks and methods to solve also increases; (4) the ’Classification Challenge’ in
which we look at how complex methods can be separated in order to provide greater reasoning
in complex classification tasks; and finally (5) the ’Curation Challenge’ where we challenge the
assumptions around bottleneck creation for the development of supervised learning methods.
Version
Open Access
Date Issued
2021-09
Online Publication Date
2022-04-26T10:32:01Z
Date Awarded
2022-03
Copyright Statement
Creative Commons Attribution NonCommercial Licence
Advisor
Kainz, Bernhard
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/S022104/1
EP/S013687/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)