This section contains the online material for the PhD thesis titled “Contributions to deep learning for imaging in radiotherapy.”. The three parts of the thesis are available directly under the following links:
According to global cancer statistics between 2012 and 2020, the annual number of new cancer cases increased from an estimated 14.1 million [36] to 19.3 million [155], while mortality rates decreased from 58.2% to 51.8%. These trends are partially attributed to advancements in early detection and treatment achieved by constantly improved medical imaging and treatment devices. In developed countries with access to the most advanced devices and procedures, the mortality rates decline the most [155].
The technological improvements of medical imaging equipment are
leading towards a growing dependence on the quality of medical data, and the increasing amount of available data is adding to the workload of the already busy medical staff. This increased burden can lead to delays and errors in patient care, highlighting the need for more efficient processes and advanced technological solutions to support the medical imaging process.
Our contributions were driven by recent advancements in two key areas: the increasing adoption of computational, automated image analysis methods in medical imaging for radiotherapy and the rapid advancement of Machine Learning (ML) solutions across virtually all scientific fields. The medical field recognized the potential of ML techniques in order to improve the accuracy and speed of medical imaging in the context of the current radiotherapy workflow [98]. The most common applications perform tasks such as segmentations [170, 63], image restoration [111], image registrations [78, 152], or modality transfer such as synthetic image generation [16, 49].
Our projects have contributed to the publicly available ML-based solutions for medical imaging tasks focusing on the radiotherapy workflow, while addressing common concerns around ML methods. This work investigated issues in the ML implementations of medical imaging applications, addressed their importance and proposed solutions. As such, our work aimed to build the relationship between medical imaging in radiotherapy and ML.
Machine learning
Over the past few years, ML has made remarkable progress and is increasingly outperforming analytical solutions in a wide range of research areas [87, 84, 46, 70]. However, these approaches are not
without their own set of challenges which must be carefully considered during their evaluations. The rate of advancement means a large portion of research is exploratory, and proposed methods and techniques quickly become widely used before proper, thorough assessment. Consequently, there has been a growing awareness of the need to critically evaluate ML solutions and to question their output predictions, as can be seen by the rise of meta-studies [96, 3], reproducibility guidelines [122], and methods for model explainability [147].
The applications of ML are numerous and varied, with the most effective ones typically being those that replace relatively simple but time-consuming manual work. Nonetheless, it is important to remain vigilant and to ensure that these techniques are used in a responsible and ethical manner, taking into account the potential societal [20] and economic [54] impacts of their deployment. With continued research and development, it is expected that ML will continue to play an increasingly important role in many fields, from healthcare through finance to manufacturing, and beyond [140].
Imaging in radiotherapy
Radiotherapy is a treatment method used on around half of all the cancer patients in Sweden [9]. Medical imaging is used for treatment planning, which ensure that the radiation is delivered precisely to the tumor while minimizing damage to healthy tissue [68]. The current radiotherapy workflow employs multiple imaging modalities which leads to labor-intensive data acquisition and processing, each modality with its own advantages together with a set of artefacts and potential issues that must be carefully considered [115].
For instance, Magnetic Resonance Imaging (MRI) is often used in radiotherapy planning due to its superior soft tissue contrast compared to Computed Tomography (CT), however the CT scans are still necessary for dose calculations [83]. This means that for improving organ delineations, an additional hardware is necessary, increasing the scan time and the work time for operating the scanners, and introduces an additional uncertainty in the case of patient motion or other anatomical mismatches between the
images [161, 131].
Automating certain image processing steps can prove beneficial in
streamlining the radiotherapy workflow, and allow medical staff to focus on tasks that require human input [19]. Despite the increasing amount of available medical data, there are still many aspects of the workflow that require manual processing. While radiotherapy treatment has made significant progress in recent years, further research is needed to develop more efficient and accurate imaging workflows that can maximize the benefits of automation while keeping the impact of potential drawbacks and uncertainties to a minimum [120].
The role of ML in radiotherapy
ML holds great promise for improving the efficiency of medical imaging tasks [157, 5]. However, in the case of medical applications, extra attention must be paid to ensure that the results are reliable, robust, and interpretable [164]. The solutions need to be validated in unfamiliar circumstances to assess how well they generalize to different situations [91, 64]. Furthermore, it is essential to document the development and testing of the models rigorously to ensure
transparency and reproducibility [109, 12]. Pre-processing is another critical step in the imaging workflow, and the outputs of ML models can be challenging to understand without sufficient ML knowledge. While the goal is to reduce the burden on clinicians, their knowledge of ML is crucial in the correct interpretation of the model outputs, particularly in the context of medical decision-making [139].
With this exhaustive list of challenges, meaningful ML solutions to
medical imaging tasks cannot be developed without a thorough understanding of both fields.
Problem specification
Automated ML solutions already exist to support the medical staff or professionals in the radiotherapy decision-making process [5, 98]. However, their inherent limitations have often not been addressed sufficiently to be considered for integration in the radiotherapy pipeline. Despite the large number of ML solutions, their generalization towards new data and other limitations need to be well established for them to be useful. Our aim in this thesis is to provide publicly available ML solutions that tackle issues of generalization and reproducibility, for medical imaging tasks that are relevant and generally useful.
Thesis structure
The presented thesis provides an examination of the two key components of our work: ML and imaging in radiotherapy. The initial sections of the thesis aim to establish a comprehensive understanding of the fundamental concepts that underpin our contributions. We begin by providing an individual description of both ML and imaging in radiotherapy, highlighting their respective roles in our work. Afterwards, we explore the current relationship between these two fields and their potential future direction.
The thesis aims to provide a valuable resource for researchers in the
field of medical imaging with ML solutions. However, the focus is on the methods and concepts that are relevant in the included papers, and these receive the most attention, while other methods that might be equally significant are mentioned less.
Supplementary material
This thesis incorporates a range of interactive figures designed to build an understanding of the fundamental principles of ML and MRI. We recognize that the concepts of ML and MRI can be complex, especially for readers unfamiliar with the topic, and our aim was to make them accessible to a broad audience. The interactive figures provide an engaging and intuitive way to explore these concepts and enhance the learning experience.