Wasserstein Generative Adversarial Networks (WGANs) can be used to generate realistic samples from complicated image distributions. The Wasserstein metric used in WGANs is based on a notion of distance between individual images, which induces a notion of distance between probability distributions of images. So far the community has considered l(2) as the underlying distance. We generalize the theory of WGAN with gradient penalty to Banach spaces, allowing practitioners to select the features to emphasize in the generator. We further discuss the effect of some particular choices of underlying norms, focusing on Sobolev norms. Finally, we demonstrate a boost in performance for an appropriate choice of norm on CIFAR-10 and CelebA.

We propose using the Wasserstein loss for training in inverse problems. In particular, we consider a learned primal-dual reconstruction scheme for ill-posed inverse problems using the Wasserstein distance as loss function in the learning. This is motivated by miss-alignments in training data, which when using standard mean squared error loss could severely degrade reconstruction quality. We prove that training with the Wasserstein loss gives a reconstruction operator that correctly compensates for miss-alignments in certain cases, whereas training with the mean squared error gives a smeared reconstruction. Moreover, we demonstrate these effects by training a reconstruction algorithm using both mean squared error and optimal transport loss for a problem in computerized tomography.

In this paper, we propose a framework for gradually improving the quality of an already existing image descriptor. The descriptor used in this paper (Afkham et al., 2013) uses the response of a series of discriminative components for summarizing each image. As we will show, this descriptor has an ideal form in which all categories become linearly separable. While, reaching this form is not feasible, we will argue how by replacing a small fraction of these components, it is possible to obtain a descriptor which is, on average, closer to this ideal form. To do so, we initially identify which components do not contribute to the quality of the descriptor and replace them with more robust components. Here, a joint feature selection method is used to find improved components. As our experiments show, this change directly reflects in the capability of the resulting descriptor in discriminating between different categories.

Karolinska University Hospital, Huddinge, Sweden, has long desired to plan hip prostheses with Computed Tomography (CT) scans instead of plain radiographs to save time and patient discomfort. This has not been possible previously as their current software is limited to prosthesis planning on traditional 2D X-ray images. The purpose of this project was therefore to create an application (software) that allows medical professionals to derive a 2D image from CT images that can be used for prosthesis planning.

In order to create the application NumPy and The Visualization Toolkit (VTK) Python code libraries were utilised and tied together with a graphical user interface library called PyQt4. The application includes a graphical interface and methods for optimizing the images for prosthesis planning.

The application was finished and serves its purpose but the quality of the images needs to be evaluated with a larger sample group. 

During the last few years, segmentation architectures based on deep learning achieved promising results. On the other hand, attention networks have been invented years back and used in different tasks but rarely used in medical applications. This thesis investigated four main attention mechanisms; Squeeze and Excitation, Dual Attention Network, Pyramid Attention Network, and Attention UNet to be used in medical image segmentation. Also, different hybrid architectures proposed by the author were tested. Methods were tested on a kidney tumor dataset and against UNet architecture as a baseline. One version of Squeeze and Excitation attention outperformed the baseline. Original Dual Attention Network and Pyramid Attention Network showed very poor performance, especially for the tumor class. Attention UNet architecture achieved close results to the baseline but not better. Two more hybrid architectures achieved better results than the baseline. The first is a modified version of Squeeze and Excitation attention. The second is a combination between Dual Attention Networks and UNet architecture. Proposed architectures outperformed the baseline by up to 3% in tumor Dice coefficient. The thesis also shows the difference between 2D architectures and their 3D counterparts. 3D architectures achieved more than 10% higher tumor Dice coefficient than 2D architectures.

Objective

The purpose of this study was to assess if ultra-low-dose CT is a useful clinical alternative to digital radiographs in the evaluation of acute wrist and ankle fractures.

Materials and methods

An ultra-low-dose protocol was designed on a 256-slice multi-detector CT. Patients from the emergency department were evaluated prospectively. After initial digital radiographs, an ultra-low-dose CT was performed. Two readers independently analyzed the images. Also, the radiation dose, examination time, and time to preliminary report was compared between digital radiographs and CT.

Results

In 207 extremities, digital radiography and ultra-low-dose CT detected 73 and 109 fractures, respectively (p < 0.001). The odds ratio for fracture detection with ultra-low-dose CT vs. digital radiography was 2.0 (95% CI, 1.4–3.0). CT detected additional fracture-related findings in 33 cases (15.9%) and confirmed or ruled out suspected fractures in 19 cases (9.2%). The mean effective dose was comparable between ultra-low-dose CT and digital radiography (0.59 ± 0.33 μSv, 95% CI 0.47–0.59 vs. 0.53 ± 0.43 μSv, 95% CI 0.54–0.64). The mean combined examination time plus time to preliminary report was shorter for ultra-low-dose CT compared to digital radiography (7.6 ± 2.5 min, 95% CI 7.1–8.1 vs. 9.8 ± 4.7 min, 95% CI 8.8–10.7) (p = 0.002). The recommended treatment changed in 34 (16.4%) extremities.

Conclusions

Ultra-low-dose CT is a useful alternative to digital radiography for imaging the peripheral skeleton in the acute setting as it detects significantly more fractures and provides additional clinically important information, at a comparable radiation dose. It also provides faster combined examination and reporting times.

Currently, quantum walk is being explored as a potential method for analyzing medical images. Taking inspiration from Grady's random walk algorithm for image processing, we have developed an approach that leverages the quantum mechanical advantages inherent in quantum walk to detect and segment medical images. Furthermore, the segmented images have been evaluated in terms of clinical relevance. Theoretically, quantum walk algorithms have the potential to offer a more efficient method for medical image analysis compared to traditional methods of image segmentation, such as classical random walk, which do not rely on quantum mechanics. Within this field, there is significant potential for development, and it is of utmost importance to continue exploring and refining these methods. However, it should be noted that there is a long way to go before this becomes something that can be applied in a clinical environment.

Today, movement measurements are made manually in healthcare using a goniometer. The measurements are often time-consuming and specially trained practitioners are needed, furthermore the readings are dependent on the practitioner's eye measurements. New technology has been introduced to make the same measurements using external sensors which have the disadvantage that they can be costly and space consuming. Recently, depth camera technology has introduced an alternative solution to make it more efficient for both patients and practitioners. The purpose of this study is to investigate whether Apple depth camera technology can replace current technology by being able to perform motion measurements with only one Ipad/Iphone with a built-in depth camera. For this, a prototype has been developed to be able to automatically calculate medically relevant angles by filming a person. Various motion measurements have been performed and the measurement results have been analysed. Apple depth camera has high precision in the measured values. However, a larger study needs to be performed in order to be able to determine whether Apple's depth camera could replace current technology.

The CPT-violation parameters Re(δ) and Im(δ) determined recently by CPLEAR are used to evaluate the K⁰– mass and decay-width differences, as given by the difference between the diagonal elements of the neutral-kaon mixing matrix (M−iΓ/2). The results – GeV and GeV – are consistent with CPT invariance. The CPT invariance is also shown to hold within a few times 10⁻³–10⁻⁴ for many of the amplitudes describing neutral-kaon decays to different final states.

We apply a forward dispersion relation to the regeneration amplitude for kaon scattering on 12" style="position: relative;" tabindex="0" id="MathJax-Element-1-Frame" class="MathJax">12C using all available data. The CPLEAR data at low energies allow the determination of the net contribution from the subthreshold region which turns out to be much smaller than earlier evaluations, solving a long standing puzzle.

Neutral-kaon decays to πeν were analysed to determine the q² dependence of the K⁰_e3 electroweak form factor f₊. Based on 365612 events, this form factor was found to have a linear dependence on q² with a slope λ₊=0.0245±0.0012_stat±0.0022_syst.

Nowadays state of the art Single Photon Emission Computed Tomography (SPECT) system employs a collimator to only allow photons incident on the detector close to a 90° angle to pass through and get detected. The source position is then derived based on the straight lines of incoming gamma rays to the detector. The efficiency of SPECT is low, in the order of 10^-6. One reason for the inefficiency is that the collimator is blocking a big fraction of the incident photons. What is now considered to be the next generation of SPECT-devices is based on Compton imaging detectors, where a collimator is not needed for image reconstruction. This thesis evaluates the performance of an interaction chain sorting algorithm for a Compton imaging detector for medical use and investigates how limits for the allowed Figure of Merit (FOM) and Rayleigh scatter impacts the accuracy and efficiency of the algorithm at different source energies. The efficiency of the algorithm applied to the simulation used in this thesis was in the order of 10^-1 which was a significant increase compared to nowadays state of the art SPECT. Based on the trade off between accuracy and efficiency, setting a limit for the allowed FOM was not desirable in order to optimise the algorithm. The impact of Rayleigh scatter on accuracy was minimal leading to the conclusion that the algorithm does not need to be modified considering Rayleigh scatter.

Magnetic particle imaging (MPI) is a rapidly developing medical imaging modality that exploits the non- linear response of magnetic nanoparticles (MNPs). Color MPI widens the functionality of MPI, empowering it with the capability to distinguish differentMNPs and/orMNP environments. The system function approach for color MPI relies on extensive calibrations that capture the differences in the harmonic responses of the MNPs. An alternative calibration-free x-space-basedmethod called TAURUS estimates amap of the relaxation time constant, tau, by recovering the underlyingmirror symmetry in the MPI signal. However, TAURUS requires a back and forth scanning of a given region, restricting its usage to slow trajectories with constant or piecewise constant focus fields (FFs). In this work, we propose a novel technique to increase the performance of TAURUS and enable tau map estimation for rapid andmultidimensional trajectories. The proposed technique is based on correcting the distortions on mirror symmetry induced by time-varying FFs. We demonstrate via simulations and experiments in our in-house MPI scanner that the proposed method successfully estimates high-fidelity tau maps for rapid trajectories that provide orders of magnitude reduction in scanning time (over 300 fold for simulations and over 8 fold for experiments) while preserving the calibration-free property of TAURUS.

Cancer is a major public health problem, accounting for an estimated 10 million deaths worldwide in 2020 alone. Rapid advances in the field of image acquisition and hardware development over the past three decades have resulted in the development of modern medical imaging modalities that can capture high-resolution anatomical, physiological, functional, and metabolic quantitative information from cancerous organs. Therefore, the applications of medical imaging have become increasingly crucial in the clinical routines of oncology, providing screening, diagnosis, treatment monitoring, and non/minimally-invasive evaluation of disease prognosis. The essential need for medical images, however, has resulted in the acquisition of a tremendous number of imaging scans. Considering the growing role of medical imaging data on one side and the challenges of manually examining such an abundance of data on the other side, the development of computerized tools to automatically or semi-automatically examine the image data has attracted considerable interest. Hence, a variety of machine learning tools have been developed for oncological image analysis, aiming to assist clinicians with repetitive tasks in their workflow.

This thesis aims to contribute to the field of oncological image analysis by proposing new ways of quantifying tumor characteristics from medical image data. Specifically, this thesis consists of six studies, the first two of which focus on introducing novel methods for tumor segmentation. The last four studies aim to develop quantitative imaging biomarkers for cancer diagnosis and prognosis.

The main objective of Study I is to develop a deep learning pipeline capable of capturing the appearance of lung pathologies, including lung tumors, and integrating this pipeline into the segmentation networks to leverage the segmentation accuracy. The proposed pipeline was tested on several comprehensive datasets, and the numerical quantifications show the superiority of the proposed prior-aware DL framework compared to the state of the art. Study II aims to address a crucial challenge faced by supervised segmentation models: dependency on the large-scale labeled dataset. In this study, an unsupervised segmentation approach is proposed based on the concept of image inpainting to segment lung and head-neck tumors in images from single and multiple modalities. The proposed autoinpainting pipeline shows great potential in synthesizing high-quality tumor-free images and outperforms a family of well-established unsupervised models in terms of segmentation accuracy.

Studies III and IV aim to automatically discriminate the benign from the malignant pulmonary nodules by analyzing the low-dose computed tomography (LDCT) scans. In Study III, a dual-pathway deep classification framework is proposed to simultaneously take into account the local intra-nodule heterogeneities and the global contextual information. Study IV seeks to compare the discriminative power of a series of carefully selected conventional radiomics methods, end-to-end Deep Learning (DL) models, and deep features-based radiomics analysis on the same dataset. The numerical analyses show the potential of fusing the learned deep features into radiomic features for boosting the classification power.

Study V focuses on the early assessment of lung tumor response to the applied treatments by proposing a novel feature set that can be interpreted physiologically. This feature set was employed to quantify the changes in the tumor characteristics from longitudinal PET-CT scans in order to predict the overall survival status of the patients two years after the last session of treatments. The discriminative power of the introduced imaging biomarkers was compared against the conventional radiomics, and the quantitative evaluations verified the superiority of the proposed feature set. Whereas Study V focuses on a binary survival prediction task, Study VI addresses the prediction of survival rate in patients diagnosed with lung and head-neck cancer by investigating the potential of spherical convolutional neural networks and comparing their performance against other types of features, including radiomics. While comparable results were achieved in intra-dataset analyses, the proposed spherical-based features show more predictive power in inter-dataset analyses.

In summary, the six studies incorporate different imaging modalities and a wide range of image processing and machine-learning techniques in the methods developed for the quantitative assessment of tumor characteristics and contribute to the essential procedures of cancer diagnosis and prognosis.

One of the major differences between medical doctor training and machine learning is that doctors are trained to recognize normal/healthy anatomy first. Knowing the healthy appearance of anatomy structures helps doctors to make better judgement when some abnormality shows up in an image. In this study, we propose a normal appearance autoencoder (NAA), that removes abnormalities from a diseased image. This autoencoder is semi-automatically trained using another partial convolutional in-paint network that is trained using healthy subjects only. The output of the autoencoder is then fed to a segmentation net in addition to the original input image, i.e. the latter gets both the diseased image and a simulated healthy image where the lesion is artificially removed. By getting access to knowledge of how the abnormal region is supposed to look, we hypothesized that the segmentation network could perform better than just being shown the original slice. We tested the proposed network on the LIDC-IDRI dataset for lung cancer detection and segmentation. The preliminary results show the NAA approach improved segmentation accuracy substantially in comparison with the conventional U-Net architecture. 

We propose a hybrid segmentation pipeline based on the autoencoders’ capability of anomaly detection. To this end, we, first, introduce a new augmentation technique to generate synthetic paired images. Gaining advantage from the paired images, we propose a Normal Appearance Autoencoder (NAA) that is able to remove tumors and thus reconstruct realistic-looking, tumor-free images. After estimating the regions where the abnormalities potentially exist, a segmentation network is guided toward the candidate region. We tested the proposed pipeline on the BraTS 2019 database. The preliminary results indicate that the proposed model improved the segmentation accuracy of brain tumor subregions compared to the U-Net model. 

Purpose: Low-Dose Computed Tomography (LDCT) is the most common imaging modality for lung cancer diagnosis. The presence of nodules in the scans does not necessarily portend lung cancer, as there is an intricate relationship between nodule characteristics and lung cancer. Therefore, benign-malignant pulmonary nodule classification at early detection is a crucial step to improve diagnosis and prolong patient survival. The aim of this study is to propose a method for predicting nodule malignancy based on deep abstract features.

Methods: To efficiently capture both intra-nodule heterogeneities and contextual information of the pulmonary nodules, a dual pathway model was developed to integrate the intra-nodule characteristics with contextual attributes. The proposed approach was implemented with both supervised and unsupervised learning schemes. A random forest model was added as a second component on top of the networks to generate the classification results. The discrimination power of the model was evaluated by calculating the Area Under the Receiver Operating Characteristic Curve (AUROC) metric. Results: Experiments on 1297 manually segmented nodules show that the integration of context and target supervised deep features have a great potential for accurate prediction, resulting in a discrimination power of 0.936 in terms of AUROC, which outperformed the classification performance of the Kaggle 2017 challenge winner.

Conclusion: Empirical results demonstrate that integrating nodule target and context images into a unified network improves the discrimination power, outperforming the conventional single pathway convolutional neural networks.

In contemporary IT industry, Multi-tasking solutions are highly regarded as optimal solutions, because hardware is equipped with multi-core CPUs.  With Multi-Core technology, CPUs run with lower frequencies while giving same or better performance as a whole system of processing. This thesis work takes advantage of multi-threading architecture in order to run different tasks under different cores such as SIP signaling and messaging to establish one or more SIP calls, capture voice, medical data, and packetize them to be streamed over internet to other SIP agents. VoIP is designed to stream voice over IP. There is inter-protocol communication and cooperation such as between the SIP, SDP, RTP, and RTCP protocols in order to establish a SIP connection and- afterwards- stream media over the internet. We use the Microsoft COM technology in order to better the C++ component design. It allows us to design and develop code once and run it anywhere on different platforms. Using VC++ helps us reduce software design time and development time. Moreover, we follow software design standards setup by software engineers’ society. VoIP technology uses protocols such as the SIP signaling protocol to locate the user agents that communicate with each other. Pjsip is a library that allows developers to extend their design with SIP capability. We use the PJSIP library in order to sign up our own developed VoIP module to a SIP server over the Internet and locate other user agents. We implement and use the already-designed iRTP protocol instead of the RTP to stream media over the Internet. Thus, we can improve RTP packet delays and improve Quality of Service (QoS). Since medical data is critical and must not be lost, the iRTP guarantees no loss of medical data. If we want to stream voice only, we would not need iRTP, because RTP is a good protocol for voice applications. Due to the increasing Internet traffic, we need to use a reliable protocol that can detect packet loss of medical data. iRTP resolves the issue and leverages QoS. This thesis work focuses on streaming medical data and medical voice-calls using VoIP, even over small bandwidths and in high traffic periods. The main contribution of this thesis is in the parallel design of iRTP and the implementation of this very design in order to be used with Multi-Core technology. We do so via multi-threading technology to speed up the streaming of medical data and medical voice-calls. According to our tests, measurements, and result analyses, the parallel design of iRTP and the multithreaded implementation on VC++ leverage performance to a level where the average decrease in delay is 71.1% when using iRTP for audio and medical data instead of the nowadays applied case of using an RTP stream for audio and multiple TCPs streams for medical data .

An automated method for the segmentation and tracking of moving vessel walls in 2D ultrasound image sequences is introduced. The method was tested on simulated and real ultrasound image sequences of the carotid artery. Tracking was achieved via a self organizing neural network known as Growing Neural Gas. This topology-preserving algorithm assigns a net of nodes connected by edges that distributes itself within the vessel walls and adapts to changes in topology with time. The movement of the nodes was analyzed to uncover the dynamics of the vessel wall. By this way, radial and longitudinal strain and strain rates have been estimated. Finally, wave intensity signals were computed from these measurements. The method proposed improves upon wave intensity wall analysis, WIWA, and opens up a possibility for easy and efficient analysis and diagnosis of vascular disease through noninvasive ultrasonic examination.

OBJECTIVES: We have developed a method using transthoracic echocardiography in establishing optimal visualization of the aortic root, to reduce the amount of contrast medium used in each patient.

BACKGROUND: During transcatheter aortic valve implantation, it is necessary to obtain an optimal fluoroscopic projection for deployment of the valve showing the aortic ostium with the three cusps aligned in the beam direction. This may require repeat aortic root angiograms at this stage of the procedure with a high amount of contrast medium with a risk of detrimental influence on renal function.

METHODS: We studied the conventional way and an echo guided way to optimize visualisation of the aortic root. Echocardiography was used initially allowing easier alignment of the image intensifier with the transducer's direction.

RESULTS: Contrast volumes, radiation/fluoroscopy exposure times, and postoperative creatinine levels were significantly less in patients having the echo-guided orientation of the optimal fluoroscopic angles compared with patients treated with the conventional approach.

CONCLUSION: We present a user-friendly echo-guided method to facilitate fluoroscopy adjustment during transcatheter aortic valve implantation. In our series, the amounts of contrast medium and radiation have been significantly reduced, with a concomitant reduction in detrimental effects on renal function in the early postoperative phase.

Vitiligo is a skin disease where the pigment cells that produce melanin die or stop functioning, which causes white patches to appear on the body. Although vitiligo is not considered a serious disease, there is a risk that something is wrong with a person's immune system. In recent years, the use of medical image processing techniques has grown, and research continues to develop new techniques for analysing and processing medical images. In many medical image classification tasks, deep convolutional neural network technology has proven its effectiveness, which means that it may also perform well in vitiligo classification. Our study uses four deep convolutional neural networks in order to classify images of vitiligo and normal skin. The architectures selected are VGG-19, ResNeXt101, InceptionResNetV2 and Inception V3. ROC and AUC metrics are used to assess each model's performance. In addition, the authors investigate the economic benefits that this technology may provide to the healthcare system and patients. To train and evaluate the CNN models, the authors used a dataset that contains 1341 images in total. Because the dataset is limited, 5-fold cross validation is also employed to improve the model's prediction. The results demonstrate that InceptionV3 achieves the best performance in the classification of vitiligo, with an AUC value of 0.9111, and InceptionResNetV2 has the lowest AUC value of 0.8560.

Purpose Our study investigates the potential benefits of incorporating prior anatomical knowledge into a deep learning (DL) method designed for the automated segmentation of lung lobes in chest CT scans. Approach We introduce an automated DL-based approach that leverages anatomical information from the lung's vascular system to guide and enhance the segmentation process. This involves utilizing a lung vessel connectivity (LVC) map, which encodes relevant lung vessel anatomical data. Our study explores the performance of three different neural network architectures within the nnU-Net framework: a standalone U-Net, a multitasking U-Net, and a cascade U-Net. Results Experimental findings suggest that the inclusion of LVC information in the DL model can lead to improved segmentation accuracy, particularly, in the challenging boundary regions of expiration chest CT volumes. Furthermore, our study demonstrates the potential for LVC to enhance the model's generalization capabilities. Finally, the method's robustness is evaluated through the segmentation of lung lobes in 10 cases of COVID-19, demonstrating its applicability in the presence of pulmonary diseases. Conclusions Incorporating prior anatomical information, such as LVC, into the DL model shows promise for enhancing segmentation performance, particularly in the boundary regions. However, the extent of this improvement has limitations, prompting further exploration of its practical applicability.

Background: Recent advancements in in situ gene expression technologies constitute a new and rapidly evolving field of transcriptomics. With the recent launch of the 10x Genomics Visium platform, such methods have started to become widely adopted. The experimental protocol is conducted on individual tissue sections collected from a larger tissue sample. The two-dimensional nature of this data requires multiple consecutive sections to be collected from the sample in order to construct a comprehensive three-dimensional map of the tissue. However, there is currently no software available that lets the user process the images, align stacked experiments, and finally visualize them together in 3D to create a holistic view of the tissue. Results: We have developed an R package named STUtility that takes 10x Genomics Visium data as input and provides features to perform standardized data transformations, alignment of multiple tissue sections, regional annotation, and visualizations of the combined data in a 3D model framework. Conclusions: STUtility lets the user process, analyze and visualize multiple samples of spatially resolved RNA sequencing and image data from the 10x Genomics Visium platform. The package builds on the Seurat framework and uses familiar APIs and well-proven analysis methods. An introduction to the software package is available at https://ludvigla.github.io/STUtility_web_site/.

Spectral imaging is the acquisition of multiple images of an object at different energy spectra. In mammography, dual-energy imaging (spectral imaging with two energy levels) has been investigated for several applications, in particular material decomposition, which allows for quantitative analysis of breast composition and quantitative contrast-enhanced imaging. Material decomposition with dual-energy imaging is based on the assumption that there are two dominant photon interaction effects that determine linear attenuation: the photoelectric effect and Compton scattering. This assumption limits the number of basis materials, i.e. the number of materials that are possible to differentiate between, to two. However, Rayleigh scattering may account for more than 10% of the linear attenuation in the mammography energy range. In this work, we show that a modified version of a scanning multi-slit spectral photon-counting mammography system is able to acquire three images at different spectra and can be used for triple-energy imaging. We further show that triple-energy imaging in combination with the efficient scatter rejection of the system enables measurement of Rayleigh scattering, which adds an additional energy dependency to the linear attenuation and enables material decomposition with three basis materials. Three available basis materials have the potential to improve virtually all applications of spectral imaging.

Real-time 3D Echocardiography (RT3DE) has been proven to be an accurate tool for left ventricular (LV) volume assessment. However, identification of the LV endocardium remains a challenging task, mainly because of the low tissue/blood contrast of the images combined with typical artifacts. Several semi and fully automatic algorithms have been proposed for segmenting the endocardium in RT3DE data in order to extract relevant clinical indices, but a systematic and fair comparison between such methods has so far been impossible due to the lack of a publicly available common database. Here, we introduce a standardized evaluation framework to reliably evaluate and compare the performance of the algorithms developed to segment the LV border in RT3DE. A database consisting of 45 multivendor cardiac ultrasound recordings acquired at different centers with corresponding reference measurements from 3 experts are made available. The algorithms from nine research groups were quantitatively evaluated and compared using the proposed online platform. The results showed that the best methods produce promising results with respect to the experts' measurements for the extraction of clinical indices, and that they offer good segmentation precision in terms of mean distance error in the context of the experts' variability range. The platform remains open for new submissions.

The goal of this master thesis is to develop a computerized method for automatic estimation of the mammographic density of mammographic images from 5 different types of mammography units.

Mammographic density is a measurement of the amount of fibroglandular tissue in a breast. This is the single most attributable risk factor for breast cancer; an accurate measurement of the mammographic density can increase the accuracy of cancer prediction in mammography. Today it is commonly estimated through visual inspection by a radiologist, which is subjective and results in inter-reader variation.

The developed method estimates the density as a ratio of #pixels-containing-dense-tissue over #pixels-containing-any-breast-tissue and also according to the BI-RADS density categories. To achieve this, each mammographic image is:

• corrected for breast thickness and normalized such that some global threshold can separate dense and non-dense tissue.
• iteratively thresholded until a good threshold is found.  This process is monitored and automatically stopped by a classifier which is trained on sample segmentations using features based on different image intensity characteristics in specified image regions.
• filtered to remove noise such as blood vessels from the segmentation.
• Finally, the ratio of dense tissue is calculated and a BI-RADS density class is assigned based on a calibrated scale (after averaging the ratings of both craniocaudal images for each patient). The calibration is based on resulting density ratio estimations of over 1300 training samples against ratings by radiologists of the same images.

The method was tested on craniocaudal images (not included in the training process) acquired with different mammography units of 703 patients which had also been rated by radiologists according to the BI-RADS density classes. The agreement with the radiologist rating in terms of Cohen’s weighted kappa is substantial (0.73). In 68% of the cases the agreement is exact, only in 1.2% of the cases the disagreement is more than 1 class.

Background and purpose Damage to the blood-brain barrier with subsequent contrast enhancement is a hallmark of glioblastoma. Non-enhancing tumor invasion into the peritumoral edema is, however, not usually visible on conventional magnetic resonance imaging. New quantitative techniques using relaxometry offer additional information about tissue properties. The aim of this study was to evaluate longitudinal relaxation R-1, transverse relaxation R-2, and proton density in the peritumoral edema in a group of patients with malignant glioma before surgery to assess whether relaxometry can detect changes not visible on conventional images. Methods In a prospective study, 24 patients with suspected malignant glioma were examined before surgery. A standard MRI protocol was used with the addition of a quantitative MR method (MAGIC), which measured R-1, R-2, and proton density. The diagnosis of malignant glioma was confirmed after biopsy/surgery. In 19 patients synthetic MR images were then created from the MAGIC scan, and ROIs were placed in the peritumoral edema to obtain the quantitative values. Dynamic susceptibility contrast perfusion was used to obtain cerebral blood volume (rCBV) data of the peritumoral edema. Voxel-based statistical analysis was performed using a mixed linear model. Results R-1, R-2, and rCBV decrease with increasing distance from the contrast-enhancing part of the tumor. There is a significant increase in R1 gradient after contrast agent injection (P<.0001). There is a heterogeneous pattern of relaxation values in the peritumoral edema adjacent to the contrast-enhancing part of the tumor. Conclusion Quantitative analysis with relaxometry of peritumoral edema in malignant gliomas detects tissue changes not visualized on conventional MR images. The finding of decreasing R-1 and R-2 means shorter relaxation times closer to the tumor, which could reflect tumor invasion into the peritumoral edema. However, these findings need to be validated in the future.

Sammanfattning

Hjärt- och kärlsjukdommar är den ledande dödsorsaken i världen. En av det vanligaste hjärt- och kärlsjukdomarna är åderförkalkning. Sjukdomen kännetecknas av förhårdning samt plackansamling i kärl och bidrar till stroke och hjärtinfarkt. Information om kärlväggens styvhet kan spela en viktig roll vid diagnostiseringen av bland annat åderförkalkning. Skjuvvågselastografi (SWE) är en noninvasiv ultraljudsbaserad metod som idag används för att mäta elasticitet och styvhet av större mjuka vävnader som lever- och bröstvävnad. Dock används inte metoden inom kärlapplikationer, då få genomgående studier har utförts på SWE för kärl. Målet med projektet är att automatisera kvantifieringen av skjuvvågshastigheten för SWE och undersöka hur automatiseringens förmåga och begränsningar beror av automatiseringsinställningar. Med verktyg erhållna från CBH (skolan för kemi, bioteknologi och hälsa) skapades ett MATLAB-program med denna förmåga. Programmet applicerades på två fantommodeller. Automatiseringsinställningarna påverkade automatiseringen av dessa modeller olika, vilket innebar att generella optimala inställningar inte kunde finnas. Optimala inställningar beror på vad automatiseringen skall undersöka.

Purpose: To develop a framework for taking the spatial frequency composition of an imaging taskinto account when determining optimal bin weight factors for photon counting energy sensitivex-ray systems. A second purpose of the investigation is to evaluate the possible improvement comparedto using pixel based weights.Methods: The Fourier based approach of imaging performance and detectability index d0 is appliedto pulse height discriminating photon counting systems. The dependency of d0 on the bin weightfactors is made explicit, taking into account both differences in signal and noise transfer characteristicsacross bins and the spatial frequency dependency of interbin correlations from reabsorbedscatter. Using a simplified model of a specific silicon detector, d0 values for a high and a low frequencyimaging task are determined for optimal weights and compared to pixel based weights.Results: The method successfully identifies bins where a large point spread function degradesdetection of high spatial frequency targets. The method is also successful in determining how todownweigh highly correlated bins. Quantitative predictions for the simplified silicon detectormodel indicate that improvements in the detectability index when applying task-based weightsinstead of pixel based weights are small for high frequency targets, but could be in excess of 10%for low frequency tasks where scatter-induced correlation otherwise degrade detectability.Conclusions: The proposed method makes the spatial frequency dependency of complex correlationstructures between bins and their effect on the system detective quantum efficiency easier toanalyze and allows optimizing bin weights for given imaging tasks. A potential increase in detectabilityof double digit percents in silicon detector systems operated at typical CT energies (100kVp) merits further evaluation on a real system. The method is noted to be of higher relevancefor silicon detectors than for cadmium (zink) telluride detectors.

In tissue characterization and cancer diagnostics, multimodal imaging has emerged as a powerful technique. Thanks to computational advances, large datasets can be exploited to discover patterns in pathologies and improve diagnosis. However, this requires efficient and scalable image retrieval methods. Cross-modality image retrieval is particularly challenging, since images of similar (or even the same) content captured by different modalities might share few common structures. We propose a new application-independent content-based image retrieval (CBIR) system for reverse (sub-)image search across modalities, which combines deep learning to generate representations (embedding the different modalities in a common space) with robust feature extraction and bag-of-words models for efficient and reliable retrieval. We illustrate its advantages through a replacement study, exploring a number of feature extractors and learned representations, as well as through comparison to recent (cross-modality) CBIR methods. For the task of (sub-)image retrieval on a (publicly available) dataset of brightfield and second harmonic generation microscopy images, the results show that our approach is superior to all tested alternatives. We discuss the shortcomings of the compared methods and observe the importance of equivariance and invariance properties of the learned representations and feature extractors in the CBIR pipeline. Code is available at: https://github.com/MIDA-group/CrossModal_ImgRetrieval.

PURPOSE:

Veno-venous ECMO is indicated in reversible life-threatening respiratory failure without life-threatening circulatory failure. Recirculation of oxygenated blood in the ECMO circuit decreases efficiency of patient oxygen delivery but is difficult to measure. We seek to identify and quantify some of the factors responsible for recirculation in a simulation model and compare with clinical data.

METHODS:

A closed-loop real-time simulation model of the cardiovascular system has been developed. ECMO is simulated with a fixed flow pump 0 to 5 l/min with various cannulation sites - 1) right atrium to inferior vena cava, 2) inferior vena cava to right atrium, and 3) superior+inferior vena cava to right atrium. Simulations are compared to data from a retrospective cohort of 11 consecutive adult veno-venous ECMO patients in our department.

RESULTS:

Recirculation increases with increasing ECMO-flow, decreases with increasing cardiac output, and is highly dependent on choice of cannulation sites. A more peripheral drainage site decreases recirculation substantially.

CONCLUSIONS:

Simulations suggest that recirculation is a significant clinical problem in veno-venous ECMO in agreement with clinical data. Due to the difficulties in measuring recirculation and interpretation of the venous oxygen saturation in the ECMO drainage blood, flow settings and cannula positioning should rather be optimized with help of arterial oxygenation parameters. Simulation may be useful in quantification and understanding of recirculation in VV-ECMO.

Background: Computer technology enables realistic simulation of cardiovascular physiology. The increasing number of clinical surgical and medical treatment options imposes a need for better understanding of patient-specific pathology and outcome prediction. Methods: A distributed lumped parameter real-time closed-loop model with 26 vascular segments, cardiac modelling with time-varying elastance functions and gradually opening and closing valves, the pericardium, intrathoracic pressure, the atrial and ventricular septum, various pathological states and including oxygen transport has been developed. Results: Model output is pressure, volume, flow and oxygen saturation from every cardiac and vascular compartment. The model produces relevant clinical output and validation of quantitative data in normal physiology and qualitative directions in simulation of pathological states show good agreement with published data. Conclusion: The results show that it is possible to build a clinically relevant real-time computer simulation model of the normal adult cardiovascular system. It is suggested that understanding qualitative interaction between physiological parameters in health and disease may be improved by using the model, although further model development and validation is needed for quantitative patient-specific outcome prediction.

Brain magnetic resonance imaging (MRI) is an imaging modality that produces detailed images of the brain without using any ionizing radiation. From a structural MRI scan, it is possible to extract morphological properties of different brain regions, such as their volume and shape. These measures can both allow a better understanding of how the brain changes due to multiple factors (e.g., environmental and pathological) and contribute to the identification of new imaging biomarkers of neurological and psychiatric diseases. The overall goal of the present thesis is to advance the knowledge on how brain MRI image processing can be effectively used to analyze and characterize brain structure.

The first two works presented in this thesis are animal studies that primarily aim to use MRI data for analyzing differences between groups of interest. In Paper I, MRI scans from wild and domestic rabbits were processed to identify structural brain differences between these two groups. Domestication was found to significantly reshape brain structure in terms of both regional gray matter volume and white matter integrity. In Paper II, rat brain MRI scans were used to train a brain age prediction model. This model was then tested on both controls and a group of rats that underwent long-term environmental enrichment and dietary restriction. This healthy lifestyle intervention was shown to significantly affect the predicted brain age trajectories by slowing the rats' aging process compared to controls. Furthermore, brain age predicted on young adult rats was found to have a significant effect on survival.

Papers III to V are human studies that propose deep learning-based methods for segmenting brain structures that can be severely affected by neurodegeneration. In particular, Papers III and IV focus on U-Net-based 2D segmentation of the corpus callosum (CC) in multiple sclerosis (MS) patients. In both studies, good segmentation accuracy was obtained and a significant correlation was found between CC area and the patient's level of cognitive and physical disability. Additionally, in Paper IV, shape analysis of the segmented CC revealed a significant association between disability and both CC thickness and bending angle. Conversely, in Paper V, a novel method for automatic segmentation of the hippocampus is proposed, which consists of embedding a statistical shape prior as context information into a U-Net-based framework. The inclusion of shape information was shown to significantly improve segmentation accuracy when testing the method on a new unseen cohort (i.e., different from the one used for training). Furthermore, good performance was observed across three different diagnostic groups (healthy controls, subjects with mild cognitive impairment and Alzheimer's patients) that were characterized by different levels of hippocampal atrophy.

In summary, the studies presented in this thesis support the great value of MRI image analysis for the advancement of neuroscientific knowledge, and their contribution is mostly two-fold. First, by applying well-established processing methods on datasets that had not yet been explored in the literature, it was possible to characterize specific brain changes and disentangle relevant problems of a clinical or biological nature. Second, a technical contribution is provided by modifying and extending already-existing brain image processing methods to achieve good performance on new datasets.

This paper aims at proposing a method to generate atlases of white matter fibers’ geometry that consider local orientation and curvature of fibers extracted from tractography data. Tractography was performed on diffusion magnetic resonance images from a set of healthy subjects and each tract was characterized voxel-wise by its curvature and Frenet–Serret frame, based on which similar tracts could be clustered separately for each voxel and each subject. Finally, the centroids of the clusters identified in all subjects were clustered to create the final atlas. The proposed clustering technique showed promising results in identifying voxel-wise distributions of curvature and orientation. Two tractography algorithms (one deterministic and one probabilistic) were tested for the present work, obtaining two different atlases. A high agreement between the two atlases was found in several brain regions. This suggests that more advanced tractography methods might only be required for some specific regions in the brain. In addition, the probabilistic approach resulted in the identification of a higher number of fiber orientations in various white matter areas, suggesting it to be more adequate for investigating complex fiber configurations in the proposed framework as compared to deterministic tractography.

Purpose: To determine if pelvic/ovarian and omental lesions of ovarian cancer can be reliably segmented on computed tomography (CT) using fully automated deep learning-based methods.

Methods: A deep learning model for the two most common disease sites of high-grade serous ovarian cancer lesions (pelvis/ovaries and omentum) was developed and compared against the well-established “no-new-Net” framework and unrevised trainee radiologist segmentations. A total of 451 CT scans collected from four different institutions were used for training (n = 276), evaluation (n = 104) and testing (n = 71) of the methods. The performance was evaluated using the Dice similarity coefficient (DSC) and compared using a Wilcoxon test.

Results: Our model outperformed no-new-Net for the pelvic/ovarian lesions in cross-validation, on the evaluation and test set by a significant margin (p values being 4 × 10–7, 3 × 10–4, 4 × 10–2, respectively), and for the omental lesions on the evaluation set (p = 1 × 10–3). Our model did not perform significantly differently in segmenting pelvic/ovarian lesions (p = 0.371) compared to a trainee radiologist. On an independent test set, the model achieved a DSC performance of 71 ± 20 (mean ± standard deviation) for pelvic/ovarian and 61 ± 24 for omental lesions.

Conclusion: Automated ovarian cancer segmentation on CT scans using deep neural networks is feasible and achieves performance close to a trainee-level radiologist for pelvic/ovarian lesions.

Relevance statement: Automated segmentation of ovarian cancer may be used by clinicians for CT-based volumetric assessments and researchers for building complex analysis pipelines.

Key points:

• The first automated approach for pelvic/ovarian and omental ovarian cancer lesion segmentation on CT images has been presented.
• Automated segmentation of ovarian cancer lesions can be comparable with manual segmentation of trainee radiologists.
• Careful hyperparameter tuning can provide models significantly outperforming strong state-of-the-art baselines. Graphical Abstract: [Figure not available: see fulltext.]

Vessel formation from pre-existing vascular networks, known as angiogenesis, is a vital physiological process only intended to occur during controlled growth and healing. It can however be initiated by tumors to during cancer progression through the ”angiogenic switch”. Understanding angiogenesis is crucial when developing targeted therapies that aim at inhibiting tumor vascularization. Zebrafish embryos are excellent models for studying angiogenesis due to their phenotypical relevance to humans, transparency, and survivability without a functioning vascular system. Imaging approaches that capture fluorescent expression in zebrafish vessels are used to study angiogenesis in vivo, but the uneven intensity distributions, image quality variations, and indistinguishable vessel boundaries make traditional image processing techniques challenging. These image analysis challenges have introduced a demand for more automatic segmentation and quantification approaches when studying the effects of targeted therapies on vessel patterning during angiogenesis. This degree project focuses on the segmentation and quantification of intersegmental vessels (ISVs) in transgenic zebrafish embryos using deep learning and morphological operations. Ground truth masks were generated for the segmentation task by tracing the relevant vessels in one of the three available batches of images. These were used to train, validate, and test the performance of a U-Net model against a fine-tuned MedSAM model. The resulting evaluation metrics showed that the U-Net model outperformed the f ine-tuned MedSAM model, leading to its selection for the quantification task. Vessel metrics of the ISVs were produced by applying morphological operations on the generated segmentations of the remaining two batches of images. Analysis of variance (ANOVA) tests showed significant differences between treatment groups across multiple vessel metrics. The results suggest that disparities in anti-angiogenesis efficacy can be detected among the investigated treatment groups, but the degree of effectiveness cannot be determined.