Malignant melanoma is the deadliest form of skin cancer. If correctly diagnosed in time, the expected five-year survival rate can increase up to 97 %. Therefore, exploring various methods for early detection can contribute with tools which can be used to improve detection of disease and finally to make sure that help is given in time. The purpose of this work was to investigate the performance and behavior of different convolutional neural network (CNN) architectures and to explore whether presegmenting clinical images would improve the prediction results on a binary classifier system. For the purposes of this paper, the two selected CNNs were Inception v3 and DenseNet201. The networks were pretrained on ImageNet and transfer learning techniques such as feature extraction and fine-tuning were used to extract the features of the training set. Batch size was varied and five-fold cross-validation was applied during training to find the optimal number of epochs for training. Evaluation was done on the ISIC test set, the PH2 dataset and a combined set of images from Karolinska University Hospital and FirstDerm, where the latter was also cropped to evaluate presegmentation. The achieved results for the ISIC test set were AUCs of 0.66 for Inception v3 and 0.71 for DenseNet201. For the PH2 test set, the AUCs were 0.82 and 0.73. The results for the Karolinska and FirstDerm set were 0.49 and 0.42. Presegmenting the latter test set resulted in AUCs of 0.58 and 0.51. In conclusion, quality of images could have a big impact on the classification performance. Batch size seems to affect the performance and could thus be an important hyperparameter to tune. Ultimately, the Inception v3 architecture seems to be less affected by different variability why selecting this architecture for a real-world clinical image application could be more suitable. However, the networks performed much worse than state of the art results in previous papers and the conclusions are based on rather inconclusive results. Therefore more research has to be done to verify the conclusions. / Malignt melanom är den dödligaste formen av hudcancer. Om en korrekt diagnos sätts tillräckligt tidigt kan den femåriga överlevnadsgraden uppgå till 97 %. Detta gör att forskningen efter metoder för tidigarelagd upptäckt kan bidra med verktyg som i sin tur kan användas till att upptäcka sjukdom och slutligen bidra till att hjälp sätts in i tid. Målet med detta arbete var att undersöka prestanda och beteende för olika faltningsbara neurala nätverk (CNN) och att undersöka ifall försegmentering av kliniska bilder kunde förbättra resultaten i ett binärt klassificeringssystem. De utvalda faltningsbara neurala nätverksarkitekturerna var Inception v3 och DenseNet201. Nätverken var förträanade på ImageNet och "Transfer-learning"-metoder som feature extraction och fine-tuning användes för att extrahera features från träningsuppsättningen. Batch size varierades och femtalig korsvalidering användes för att hitta det optimala antalet träningsepoker. Utvärderingen gjordes med bilder i testset från ISIC, PH2 och Karolinska och FirstDerm. Bilderna i den senare datamängden beskärdes för att utvärdera försegmenteringen av kliniska bilder. De uppnådda resultaten för ISIC testmängden var AUC-värden på 0.66 för Inception v3 och 0.71 för DenseNet201. För PH2 låg AUC-värdena på 0.82 respektive 0.73. Resultaten för testmängden med bilder frön Karolinska och FirstDerm var 0.40 och 0.42. Försegmenteringen av den sistnämnda testmängden gav AUC-värden på 0.58 och 0.51. Sammanfattningsvis kan bildkvalitet ha en stor inverkan på ett nätverks klassificeringsprestanda. Batch size verkar också påverka resultaten ochkan därför vara en viktig hyperparameter att stämma. Slutligen verkar Inception v3 vara mindre känslig för olika typer av variabiltet vilket görvalet av denna arkitektur mer lämplig ifall en riktig applikation ska byggas för detektion av exempelvis kliniska bilder. Det som bör understrykas i detta arbete är att resultaten var mycket sämre än det som bäst uppvisats i föregående rapporter och att slutasatserna är baserade på relativt ickeövertygande värden. Därför efterkrävs mer forskning för att styrka slutsatserna.

Magnetic resonance imaging (MRI) is a powerful imaging tool for diagnostics, which AMRA uses to segment and quantify certain anatomical regions. This thesis investigate the possibilities of using deep learning for the particular task of AMRAs segmentation, both for ordinary regions (fat and muscle regions) and injured muscles.The main approach performs muscle and fat segmentation separately, and compares results for three approaches; a full resolution approach, a down-sample approach (trained on down-sampled images) and an ensemble approach (uses voting among the 7 best networks).The results shows that deep learning segmentation is possible for the task, with satisfactory results. The down-sampled approach works best for fat segmentation, which can be related to the inconsistently over-segmented ground truth fat masks. It is therefore unnecessary with the additional resolution, which might only impair the performance. The down-sampled approach achieves better results also for muscle segmentation. Ensemble learning does in general not improve the neither the segmentation dice score nor the biomarker predictions. Injured muscles are more difficult to predict due to smaller muscles in the particular used dataset, and an increased data versatility. As a summary, deep learning shows great potential for the task. The results are overall satisfactory (mostly for a down-sampled approach), but further work needs to be done for injured muscles in order to make it clinically useful.

Hip fractures in older adults have severe effects on patients’ morbidity as well as mortality, so it is crucial to avoid this injury through the early identification of patients at high risk. Currently, the diagnosis of osteoporosis and consequently hip fracture risk is done through the measurement of bone mineral density by a dual-energy X-ray absorptiometry (DXA) scan. However, studies show that this method is not accurate enough, and a high percentage of patients who sustain a hip fracture had non-osteoporotic DXA scans less than a year before the incidence. In this research, to enhance the hip fracture risk prediction, the effect of a femur’s geometry and bone mineral density distribution was considered in the hip fracture risk estimation. This was done through 2D and 3D statistical shape and appearance modeling of the proximal femur using standard clinical DXA scans. To assess the proposed techniques, destructive mechanical tests were performed on 16 isolated cadaveric femurs. Also, through collaboration with the Canadian Osteoporosis Study (CaMos), the proposed statistical techniques to predict the hip fracture risk were evaluated in a clinical population as well. The results of this study showed that new techniques can enhance hip fracture risk estimation; in the clinical study, 2D and 3D statistical modeling were able to improve identifying patients at high risk by 40% and 44% over the clinical standard method. Also, the percentage of correct predictions using 2D statistical models did not differ significantly from the 3D predictions. Therefore, by applying these techniques in clinical practice it could be possible to identify patients at high risk of sustaining a hip fracture more accurately and eventually reduce the incidence of hip fractures and the pain and social and economic burden that comes with it. / Thesis / Doctor of Philosophy (PhD) / Diagnosis of osteoporosis and consequently hip fracture risk is based on the measurement of bone mineral density in clinical imaging called DXA scanning. However, studies have shown that this method is not sufficient in identifying all patients at high risk of sustaining a hip fracture. The purpose of this work was to incorporate the geometry and bone mineral density distribution of the proximal femur in hip fracture risk prediction through image processing of DXA scans. Two algorithms of 2D and 3D statistical shape and appearance modeling were implemented and evaluated in a cadaveric study (comparing the predicted fracture load to measured ones) as well as a clinical study (comparing the fracture predictions to the fracture history of patients). The results indicated that new techniques can enhance hip fracture risk estimation compared to the clinical standard method, and hence the devastating injury can be prevented through applying protective measures.

Mammography screenings are performed regularly on women in order to detect early signs of breast cancer, which is the most common form of cancer. During an exam, X-ray images (called mammograms) are taken from two different angles and reviewed by a radiologist. If they find a suspicious lesion in one of the views, they confirm it by finding the corresponding region in the other view. Finding the corresponding region is a non-trivial task, due to the different image projections of the breast and different angles of compression needed during the exam. This thesis explores the possibility of using deep learning, a data-driven approach, to solve the corresponding regions problem. Specifically, a convolutional neural network (CNN) called U-net is developed and trained on scanned mammograms, and evaluated on both scanned and digital mammograms. A model based method called the arc model is developed for comparison. Results show that the best U-net produced better results than the arc model on all evaluated metrics, and succeeded in finding the corresponding area 83.9% of times, compared to 72.6%. Generalization to digital images was excellent, achieving an even higher score of 87.6%, compared to 83.5% for the arc model.

The objective of this PhD research was the development of novel 3D interactive medical platforms for medical image analysis, simulation and visualisation, with a focus on oncology images to support clinicians in managing the increasing amount of data provided by several medical image modalities. DoctorEye and Automatic Tumour Detector platforms were developed through constant interaction and feedback from expert clinicians, integrating a number of innovations in algorithms and methods, concerning image handling, segmentation, annotation, visualisation and plug-in technologies. DoctorEye is already being used in a related tumour modelling EC project (ContraCancrum) and offers several robust algorithms and tools for fast annotation, 3D visualisation and measurements to assist the clinician in better understanding the pathology of the brain area and define the treatment. It is free to use upon request and offers a user friendly environment for clinicians as it simplifies the implementation of complex algorithms and methods. It integrates a sophisticated, simple-to-use plug-in technology allowing researchers to add algorithms and methods (e.g. tumour growth and simulation algorithms for improving therapy planning) and interactively check the results. Apart from diagnostic and research purposes, it supports clinical training as it allows an expert clinician to evaluate a clinical delineation by different clinical users. The Automatic Tumour Detector focuses on abdominal images, which are more complex than those of the brain. It supports full automatic 3D detection of kidney pathology in real-time as well as 3D advanced visualisation and measurements. This is achieved through an innovative method implementing Templates. They contain rules and parameters for the Automatic Recognition Framework defined interactively by engineers based on clinicians’ 3D Golden Standard models. The Templates enable the automatic detection of kidneys and their possible abnormalities (tumours, stones and cysts). The system also supports the transmission of these Templates to another expert for a second opinion. Future versions of the proposed platforms could integrate even more sophisticated algorithms and tools and offer fully computer-aided identification of a variety of other organs and their dysfunctions.

The severity of traumatic brain injury (TBI) is known to be characterized by the shift of the middle line in brain as the ventricular system often changes in size and position, depending on the location of the original injury. In this thesis, the focus is given to processing of the CT (Computer Tomography) brain images to automatically calculate midline shift in pathological cases and use it to predict Intracranial Pressure (ICP). The midline shift measurement can be divided into three steps. First the ideal midline of the brain, i.e., the midline before injury, is found via a hierarchical search based on skull symmetry and tissue features. Second, the ventricular system is segmented from the brain CT slices. Third, the actual midline is estimated from the deformed ventricles by shape matching method. The horizontal shift in the ventricles is then calculated based on the ideal midline and the actual midline in TBI CT images. The proposed method presents accurate detection of the ideal midline using anatomical features in the skull, accurate segmentation of ventricles for actual midline estimation using the information of anatomical features with a spatial template derived from a magnetic resonance imaging (MRI) scan, and an accurate estimation of the actual midline based on the robust proposed multiple regions shape matching algorithm. After the midline shift is successively measured, features including midline shift, texture information of CT images, as well as other demographic information are used to predict ICP. Machine learning algorithms are used to model the relation between the ICP and the extracted features. By using systematic feature selection and parameter selection of the learning model, promising results on ICP prediction are achieved. The prediction results also indicate the reliability of the proposed midline shift estimation.

MRI is used to produce images of tissue in the body. DTI, specifically, makes it possible to track the effects of nerves where they are in the brain. This project includes a shell script and a guide for using the FMRIB Software Library, followed by StarTrack and then Trackvis in order to track difficult areas in the brain. The focus is on the trigeminal nerve (CN V). The method can be used to compare nerves in the same patient, or as a comparison to a healthy brain.

Este trabalho apresenta como parte de resultados, uma ampla pesquisa que permitiu identificar os grupos de pesquisas mais importantes do mundo, os quais possuem em comum o processamento de imagens médicas, mais especificamente o processamento de imagens que busca a identificação de microcalcificações mamárias. O vasto levantamento, a seleção e organização culminou na reunião de mais de cem artigos, publicados nos mais importantes periódicos da área, que mostram claramente as formas utilizadas pelos grupos de pesquisa para apresentação dos resultados encontrados pelos seus algoritmos. Esses resultados devem auxiliar o médico no diagnóstico do câncer de mama. Demonstramos neste trabalho porque as técnicas utilizadas para apresentação dos resultados são insatisfatórias e propusemos um novo método de avaliação desses resultados. O método proposto no trabalho baseia-se no teste do X&sup2 (Qui-Quadrado), nas curvas ROC (Receiver Operating Characteristic) e no teste de concordância, que juntos permitem apresentar de forma clara e objetiva as relações entre verdadeiros positivos e falsos positivos, verdadeiros negativos e falsos negativos, sensibilidade e especificidade do algoritmo analisado. O novo método é preciso e tem bases estatísticas conhecidas pelos médicos e pelos pesquisadores, facilitando sua aceitação. / This work presents as part of results, a wide investigation that it allowed to identify the principal research groups of the world, which possess in common the processing of medical images, more specifically the processing of images that search for the identification of mammary microcalcifications. The vast collection, selection and organization culminated in the meeting of more than a hundred articles, published in the most important newspapers of the area, that show the forms used by the research groups to present the results found clearly by its algorithms. Those results should assist the doctor in the diagnosis of the breast cancer. We demonstrated in this work that the techniques used for presentation of the results are unsatisfactory and we proposed a new method of evaluation of those results. The proposed method bases on the test of the X&sup2 (Qui-square), in ROC curve (Receiver Operating Characteristic) and in the agreement test, that take together allow to present in a clear and objective way the relationships among true positive and false positive, true negative and false negative, sensibility and specificity of the analyzed algorithm. The new method is precise and has statistical bases known by the clinicians and researchers, facilitating its acceptance.

A personalized cardiovascular lumped parameter model of the left-sided heart and thesystemic circulation has been developed by the cardiovascular medicine research groupat Linköping University. It provides information about hemodynamics, some of whichcould otherwise only have been retrieved by invasive measurements. The framework forpersonalizing the model is made using 4D Flow MRI data, containing volumes describinganatomy and velocities in three directions. Thus far, the inputs to this model have beengenerated manually for each subject. This is a slow and tedious process, unpractical touse clinically, and unfeasible for many subjects.This project aims to develop a tool to calculate the inputs and run the model for mul-tiple subjects in an automatic way. It has its basis in 4D Flow MRI data sets segmentedto identify the locations of left atrium (LA), left ventricle (LV), and aorta, along with thecorresponding structures on the right side.The process of making this tool started by calculation of the inputs. Planes were placedin the relevant positions, at the mitral valve, aortic valve (AV) and in the ascending aortaupstream the brachiocephalic branches, and flow rates were calculated through them. TheAV plane was used to calculate effective orifice area of AV and aortic cross-sectional area,while the LV end systolic and end diastolic volumes were extracted form the segmentation.The tool was evaluated by comparison with manually created inputs and outputs,using 9 healthy volunteers and one patient deemed to have normal left ventricular func-tion. The patient was chosen from a subject group diagnosed with chronic ischemic heartdisease, and/or a history of angina, together with fulfillment of the high risk score ofcardiovascular diseases of the European Society of Cardiology. This data was evaluatedusing coefficient of variation, Bland-Altman plots and sum squared error. The tool wasalso evaluated visually on some subjects with pathologies of interest.This project shows that it is possible to calculate inputs fully automatically fromsegmented 4D Flow MRI and run the cardiovascular avatar in an automatic way, withoutuser interaction. The method developed seems to be in good to moderate agreement withthose obtained manually, and could be the basis for further development of the model.

Digital medical atlases can contain anatomical information which is valuable for medical doctors in diagnosing and treating illnesses. The increased availability of such atlases has created an interest for computer algorithms which are capable of integrating such atlas information into patient specific dataprocessing. The field of medical image registration aim at calculating how to match one medical image to another. Here the atlas information could give important hints of which kinds of motion are plausible in different locations of the anatomy. Being able to incorporate such atlas specific information could potentially improve the matching of images and plausibility of image registration - ultimately providing a more correct information on which to base health care diagnosis and treatment decisions. In this licentiate thesis a generic signal processing framework is derived : Global Linear Optimization (GLO). The power of the GLO framework is first demonstrated quantitatively in a very high performing image denoiser. Important proofs of concepts are then made deriving and implementing three important capabilities regarding adaptive filtering of vector fields in medica limage registration: Global regularization with local anisotropic certainty metric. Allowing sliding motion along organ and tissue boundaries. Enforcing an incompressible motion in specific areas or volumes. In the three publications included in this thesis, the GLO framework is shown to be able to incorporate one each of these capabilities. In the third and final paper a demonstration is made how to integrate more and more of the capabilities above into the same GLO to perform adaptive processing on relevant clinical data. It is shown how each added capability improves the result of the image registration. In the end of the thesis there is a discussion which highlights the advantage of the contributions made as compared to previous methods in the scientific literature. / Dynamic Context Atlases for Image Denoising and Patient Safety