Thermal ablation is an effective treatment for hepatocellular carcinoma (HCC), the most common type of primary liver cancer. However, early recurrence (ER) after treatment for HCC remains a significant challenge. This study aimed to (1) establish a robust and reproducible pipeline for prognostic model development, and (2) apply this pipeline to develop a preliminary model predicting ER in a Western HCC-cohort. Clinical data and contrast-enhanced pre- and post-ablation CT-scans from the Dutch multicentre PROMETHEUS cohort were used. Predictors included clinical characteristics, radiomics predictors, deep learning (DL) based predictors from two pre-trained models (ResNet18 and VGG16), and the minimal ablation margin (MAM). Cox regression models were constructed using principal component analysis (PCA) and regularised predictor selection (LASSO). Model performance was internally validated using bootstrapping and evaluated with the concordance index (C-index), calibration plots, and decision curve analysis. Using the developed pipeline, several individual and combination models were developed. The highest C-index in the validation sets was achieved by the model incorporating radiomics predictors and the MAM, which was trained on data from 68 patients (C-index = 0.63, 95% CI: 0.44 – 0.79). Despite promising performance in training sets, all models showed signs of overfitting and limited calibration in unseen data. The pipeline adheres to current guidelines and supports further refinement and validation. An externally validated prognostic model could help personalise treatment and follow-up strategies after thermal ablation in western HCC patients.

With a 5-year survival rate of only 9%, pancreatic cancer is one of the deadliest types of cancer. Among other things, this is caused by the extreme difficulty of diagnosing recurrent pancreatic cancer in an early stage. One of the important next steps in discovering pancreatic cancer automatically on a CT scan is finding healthy pancreatic tissue. This research explores the feasibility of segmenting the pancreas on a CT scan using a deep learning approach, focusing on postoperative cases after pancre- atic resection, by combining state-of-the-art segmentation models for the preoperative pancreas with novel techniques, such as 3D Multi-Scale Convolutional Blocks (MCBs) and KNet architectures. Us- ing pretraining on preoperative data, a complete segmentation pipeline was designed to segment the pancreas in the postoperative state. The experimental results demonstrate that deep learning models can effectively segment the pancreas, despite the increased complexity and difficulty of the postoper- ative state. Notably, employing advanced techniques such as the KNet architecture and MCBs yields significant performance improvements. The newly designed architectures presented in this research, MKNet, MSKNet, and MAKNet, achieve state-of-the-art results for both preoperative and postoperative pancreas segmentation, greatly improving the Hausdorff Distance (HD) and 95th Percentile Hausdorff Distance (HD95) in particular. In the preoperative state, substantial performance improvements were observed compared to the previous state-of-the-art, with a 12.17% increase for HD and a 14.64% in- crease for HD95. Similarly, in the postoperative state, improvements of 11.99% for HD and 13.25% for HD95 were measured. Additionally, qualitative evaluations were conducted with seven experienced radiologists and radiotherapists to validate the algorithm’s performance in a clinical context. Remark- ably, in 83% of the cases, the algorithm was evaluated to accurately segment the pancreas, requiring minimal or no modifications according to the medical experts. This study not only contributes to the state-of-the-art in pancreas segmentation but also introduces comprehensive quantitative and qualitative evaluation methodologies. Furthermore, it addresses the segmentation of the postoperative pancreas, which has not yet been touched upon in previous re- search. Limitations are acknowledged, such as the availability of a limited dataset, and future research directions are proposed, including generalizability and enhancements to the proposed models. Despite the limited availability of data, which impacts the performance and generalization capabilities to some degree, the research in this thesis showcases the capability of the MKNet family architectures to ac- curately segment the postoperative pancreas, offering potential benefits for medical applications, data annotation acceleration, and future research in this domain.

Coronary artery disease (CAD) is one of the leading causes of death and disability worldwide. In CAD, the coronary arteries, that supply the myocardium with oxygen, are narrowed or even blocked by a process called atherosclerosis. Invasive coronary angiography (ICA) is the gold standard for the diagnosis of CAD, as well as for intraprocedural guidance of percutaneous coronary interventions. At this moment, stenosis severity is determined via visual inspection by a cardiologist. This method has several important drawbacks: a significant inter- and intra-rater variability and a high positive prediction bias. Currently available additional assessment techniques improve results, however, they come with prolonged procedural time, complication risks and increased costs. Additionally, a relevant number of syndromes exists that cannot be diagnosed sufficiently with these techniques. Automated software that uses all information within the ICA images and relates it to the context of complaints and outcomes, could be a valuable tool for improvement. Since machine learning (ML) can find relations between patient groups based on images, this could be a possible solution. However, despite promising results, clinical implementation of ML is still limited. Limited clinical applicability of developed algorithms and a lack of large high-quality datasets are the cause of this. In this research it was investigated if ML can help in solving the diagnostic and therapeutic challenges encountered in interventional cardiology (IC) and what is needed to apply ML to ICA images and other sources of coupled medical data. This was done in several steps. First the expectations and perceived barriers by interventional cardiologists on ML-based algorithms in clinical practice were assessed. Next, the feasibility of health insurance code-based querying of electronic health records (EHRs) for the creation of ML datasets was assessed. Third, a proof-of-concept study on creating a dataset and a deep learning network for predicting lesion significance on ICA images was carried out. Last, a roadmap for the curation of data for the development of ML models in IC was created.
Generally, interventional cardiologists have positive expectations for applying ML in their clinical practice. Furthermore, the willingness to collaborate in the development and clinical validation of ML algorithms is high. This is essential for translating ML models to clinical practice. Health insurance code-based querying of EHRs is not a feasible approach for creating datasets for the complex syndromes which currently pose diagnostic challenges in IC. However, the EHRs hold valuable data for the development of ML datasets. The same goes for the data collection strategy in the proof-of-concept study and it was shown that it is feasible to train deep learning networks on this data.
ML shows promising results for solving the diagnostic and therapeutic challenges encountered in interventional cardiology and directions for further research were identified. For the creation of algorithms that can be applied in clinical practice, close collaboration between ML professionals and clinicians is needed. Besides this, further research is needed to develop scalable strategies for the creation of large datasets, containing adequately labelled patients that represent the real-life population.