Background:
Interactive segmentation models combine auto-segmentation methods with user interaction to overcome the inconvenience of manually adjusting contours generated by imperfect auto-contouring models. However, these models have not yet been implemented for tumor target volume segmentation in clinical radiotherapy settings. Therefore, this study validates a previously developed auto-contour refinement tool at the LUMC for Head-and-Neck (H&N) radiotherapy, demonstrating its robustness and trustworthiness.

Methods:
A user study with six non-expert participants was performed, who iteratively refined a contour-refinement model prediction to align as closely as possible with the corresponding ground truth for six tumor volumes from six patients. The contour-refinement model updated its prior predictions based on user-provided foreground (tumor) and background (non-tumor) scribbles. This enabled Three Dimensional (3D) refinement until a satisfactory result was achieved.
User inputs were collected and evaluated using performance metrics such as Dice and Surface Dice to evaluate robustness of the model, along with two newly introduced evaluation metrics proposed in this study to evaluate trustworthiness: local and non-local (Surface) Dice.

Results:
Robust behavior is observed, as the model reacts in a highly consistent manner across all users. Only minor differences in model performance (Delta Dice scores of 0.1407 vs. 0.1296) were observed across users when different user inputs were applied.

The AI pencil yields a strong initial improvement compared to manual annotations (27.4% vs. 6.4%, Wilcoxon p = 0.047), whereas subsequent iterations show variability. This variability was frequently observed in cases of incorrect user input, distortions caused by dental implants, anatomically complex regions, and during the segmentation of slices at the tumor boundaries.
In all other cases the model showed a high trustworthiness, as it follows the users intent during the contouring process.

Conclusion:
The incorporation of user feedback into the contour-refinement model results in a rapid improvement in segmentation quality across the entire volume. However, manual refinement by clinicians remains necessary for anatomically complex slices.
Overall, this research shows that the model is robust to variations in user input and (apart from the first few iterations) there are no spurious changes in non-local areas. These are important findings when working towards clinical adoption of these interactive contour refinement models. ... Read more

Background. Accurate prediction of ischemic lesion volume (ILV) in the subacute phase is essential to estimate functional outcome, as the two are positively associated. Ischemic lesions can continue to evolve between 24 hours and 1 week after stroke onset, even after successful treatment. Radiomics offers a promising approach for ILV prediction using non-contrast computed tomography (NCCT), the first-line imaging modality in AIS. However, applying CT radiomics in AIS remains challenging, as ischemic lesion segmentation is time consuming and challenging, due to its low contrast. 
Objective. This study aims to investigate whether radiomic features extracted from post-treatment NCCT scans, acquired at 24 hours after stroke onset, can be used to predict the subacute ischemic lesion volume at 1 week. In addition, it explores whether simplified annotations are feasible for radiomic feature extraction. As a secondary analysis, this study explores whether incorporating clinical data has an added value for this prediction task. 
Methods. Patients from the MR CLEAN-NOIV trial, with 24-hour and 1-week follow-up NCCT scans available, were included. The included patients were randomly divided into a pre-training set (80%) and a test set (20%). Radiomic features were extracted from the 24-hour NCCT scan using three annotation types: (1) the original segmentation, (2) a bounding box annotation, and (3) a circle annotation. Feature selection included reproducibility filtering, low-variance filtering, correlation-based clustering, and Least Absolute Shrinkage and Selection Operator (LASSO) regression. Three XGBoost radiomics regression models were trained, using five-fold cross-validation. Additionally, three combined models, using a combination of clinical and radiomic features, and two clinical models, using only clinical features, were constructed. The performance of the models was evaluated on the test set using the coefficient of determination (R²), concordance correlation coefficient (CCC), mean absolute error (MAE), and root mean squared error (RMSE). Feature importance was assessed using SHapley Additive exPlanations (SHAP). 
Results. The radiomics model based on the original segmentation achieved a high predictive performance (R² = 0.89, CCC = 0.95, MAE = 24 mL, RMSE = 31 mL). The radiomics model based on the bounding box achieved comparable performance, and the model based on the circle annotation yielded significantly lower performance. Incorporating clinical features did not significantly improve the predictive performance of the radiomics models. Across all well-performing models including radiomic features, the Run Length Non-Uniformity radiomic feature was a strong predictor of the 1-week ILV. 
Conclusion. Radiomic features extracted from 24-hour NCCT scans can accurately predict the subacute ILV at 1-week. A simplified bounding box annotation is a simpler and effective alternative to the detailed lesion segmentation, whereas the circle annotation showed poor performance and is not a good alternative for radiomic feature extraction in this context. These findings demonstrate the potential of radiomics and the use of simplified annotations for feature extraction to predict patient prognosis and guide personalized stroke care. However, further research is required before these models can be considered for clinical use. ... Read more

Background: Real-time adaptive radiotherapy workflows require fast spatial dose calculations with clinical accuracy. Modern physics-based dose calculation algorithms often compromise between speed and accuracy. In contrast, deep learning methods have shown to be effective at predicting spatial dose distributions with high accuracy in sub-second times. Very High Energy Electrons (VHEE) have shown potential as a treatment modality in recent years due to their penetrative ability and conformality. Creating a need for fast and accurate VHEE spatial dose calculations.
Purpose: This study presents a deep learning-based algorithm that utilizes convolutional layers and the self-attention mechanism to predict VHEE beam spatial dose distributions in sub-second times.
Methods: The presented Electron-Dose Transformer Algorithm (E-DoTA) maps the 3D patient geometry and beam characteristics (using a vector representing the beam energy and a 3D Gaussian with a specific Gaussian positional spread representing the beam shape) to
a 3D dose distribution. E-DoTA uses a series of 3D convolutional layers to extract features from the patient geometries and beam shape, followed by a transformer to route information between the extracted features and the added energy vector, which are then upsampled to
the 3D dose distribution. E-DoTA is trained on 60,000 combinations of patient geometries and beam characteristics, derived from 15,000 independent patient geometries based on 12 distinct patient CT scans from the abdomen region. The model’s accuracy and prediction
speed are assessed using 8,000 previously unseen patient geometries and beam characteristics.
Results: E-DoTA predicts dose distributions of VHEE beams with high accuracy, achieving a gamma pass rate of 97.05% ± 3% (3mm, 1%) and an average relative dose error of 0.254% ± 0.096% in approximately 131 ms.
Conclusions: The fast and high-accuracy dose predictions allow the speed-up of VHEE spatial dose distribution calculations, which is currently only provided by slow Monte Carlo algorithms. Further optimizations to E-DoTA could allow for more accurate and faster dose calculations, thereby potentially accelerating VHEE radiotherapy workflows. ... Read more

Ischemic stroke, a leading cause of death and disability worldwide, occurs when a blood vessel is occluded by a thrombus. Current therapies for ischemic stroke, include Intravenous thrombolysis (IVT) and Endovascular thrombectomy (EVT). EVT relies on a Thrombolysis in Cerebral Infarction (TICI) score for assessing treatment effectiveness. This score, based on the visual evaluation of medical images by physicians, suffers from inter- and intra-observer variability, as it is influenced by the individual rater’s judgment. Digital Subtraction Angiography (DSA) imaging is commonly utilized both before therapy to identify the occluded vessel and after therapy to evaluate the treatment outcome. Accurate vessel correspondence before and after treatment is crucial for a reliable assessment. To enhance current evaluation methods and address this challenging task, we propose two automated approaches for determining vessel correspondence in pre- and post-EVT DSA imaging. The proposed methods
utilize graphical representations of the cerebral vascular network and distinct matching procedures. We refer to the methods as registration-based vessel matching (RB-VM) and graph based vessel matching (GB-VM). The methods were evaluated using manually annotated data with the RB-VM and GB-VM methods achieving a recall of 82.7% (78.2; 85.7) and 51.3% (47.4; 54.4) respectively. This work marks a significant step towards automatic stroke therapy assessment and showcases the potential benefits of graph based algorithms for this task, paving the way for more reliable and objective treatment assessments. ... Read more

In an attempt to find alternatives for solving partial differential equations (PDEs)
with traditional numerical methods, a new field has emerged which incorporates
the residual of a PDE into the loss function of an Artificial Neural Network. This
method is called Physics-Informed Neural Network (PINN). In this thesis, we study dense neural networks (DNNs), including codes developed in the context of this bachelor project. We derive the backpropagation equations necessary for training and use different configurations in a DNN to test its interpolating accuracy. We distinguish between a-PINNs which use automatic differentiation to evaluate a PDE, and n-PINNs which approximate differential operators in a PDE with numerical differentiation. We compare both PINNs on the harmonic oscillator, the 1D heat equation and the 1-soliton and 2-soliton solutions of the Korteweg-De Vries (KdV) equation. Both PINNs could accurately converge to the solution, except to the 2-soliton solution, where the a-PINN outperformed the n-PINN. Furthermore, we tested a highly nonlinear problem of the KdV equation, which can be described by a train of solitons. We observed that PINNs are inaccurate if insufficient training samples are used for training. Adding training samples on the interior from a numerical solution leads to a good qualitative agreement, though more effort is required to find a better network configuration to obtain more accurate predictions.
Additionally, PINNs were used for inverse problems to derive an unknown coefficient in a PDE and proved to be highly accurate for noiseless data. When we
generated training samples with 10% noise from a uniform distribution, the PINN
results’ relative error stayed within a margin of under 2%. However, inverse PINNs are much more inefficient compared to nonlinear least squares methods like the Levenberg–Marquardt algorithm.
As of now, PINNs are still very early in development and stand no match against
traditional numerical methods to a known PDE. They may, however, provide a
useful alternative in the future as they are constantly being improved. ... Read more

One of the principal signs of disease progression in brain tumor patients is an increase in tumor size between time-separated medical image acquisitions. The current diagnosis of tumor progression is based on visual appraisal or manual measurement of largest diameters, neither of which is a fully quantitative measure. The RANO criteria dictate that more than 25% growth in the product of two largest diameters indicates tumor progression. A more accurate assessment can be done if the tumor is fully segmented and a measurement of tumor volume is obtained. Tumor volume measurements have been shown to reduce variation due to inter- and intra-rater variability, patient position in the scanner and subjective determination of the largest diameter. Both tumor segmentation and size computation must currently be done manually and can be very time-consuming. A number of automated algorithms have been developed for lesion segmentation, but none have yet made it into clinical practice. This project focuses on the gap in currently available tools, which is the lack of volumetric analysis, particularly as it pertains to longitudinal data. We propose a tool that offers intuitive, flexible, and easy to evaluate quantitative analysis of uploaded tumor segmentation data. The tool computes volume metrics, such as tumor volume and dimensions, and visualizes them in a statistical analysis. It is based on R and utilizes modern packages for data analysis, and is deployed as a web interface using the Shiny R package. It enables the user to upload their own lesion segmentation data in NIfTI format. For easy demonstration and proof of concept, it makes use of default data from BraTS, a publicly shared repository of brain lesion data. This tool can be used for segmentation exploration of groups of patients or study participants, and it scales to any cohort size. The tool can be accessed via shinyapps.io [1] [2] and the code is available at my GitHub page [3]. ... Read more

The Recurrent Inference Machine (RIM) has been developed as an alternative to the clinically used Compressed Sensing (CS) algorithm, using Deep Learning (DL). A common issue with DL networks is the generalization of the network to features that have not been trained for. In this study we evaluate the robustness of the RIM to white matter lesions in FLuid Attenuated Inversion Recovery (FLAIR) brain MRI data. We are evaluating two pre-trained RIM networks, one trained on T1 brain data and another trained on T2 knee data. This evaluation was done by comparing the two networks to CS in terms of the average relative Signal-to-Noise Ratio (SNR) and Contrast Resolution (CR) that is achieved on 15 datasets acquired from Multiple Sclerosis patients. From these comparisons it shows that the network trained on T2 knee data performs similar to CS in terms of the relative SNR, while having a higher CR. The network trained on T1 brain data has both a lower relative SNR and CR, compared to CS. The data suggest that the RIM trained on T2 knee data is robust to the inclusion of lesions in an area that the network was not trained on. ... Read more

This project aims to recreate intensity patterns using Fraunhofer diffraction as a means of simulation. These intensity patterns are created by phase shifting specific parts of an incoming field of light. These phase shifts are determined by a B-spline surface, which is in turn controlled by so-called control points. Only a handful of control points can describe a whole surface. The position of these control points is then determined using machine learning and specifically a technique inspired by ‘physics-informed neural networks’, which were introduced last year by Raissi et al. [1]. With this method, simple experiments which sought to recreate intensity patterns known to be in the solution space were carried out. These experiments showed some success, but suffered from the fact that they used too sensitive parameters in the input of the machine learning model, reducing the sophisticated method to a Monte Carlo search, or they used no input at all, which degraded the machine learning model to simple parameter optimization. Nevertheless, these experiments showed that this method has the potential to be used in more flexible optical setups, where multiple configurations can yield the same intensity pattern or where changing the parameters defining the setup do not induce enormous changes in the resulting intensity pattern. In addition, the proposed method relies upon Fraunhofer diffraction, which, when discretized for numerical computation, introduces aliasing issues when the incoming field changes too rapidly. This phenomenon was especially apparent when using point sources that create spherical wave fronts. A possible solution for this issue is to consider ray tracing techniques in future research. ... Read more