No description available.

Convolutional neural networks are equivariant to translations; equivariance to other symmetries, however, is not defined and the class output may vary depending on the input's orientation. To mitigate this, the training data can be augmented at the cost of increased redundancy in the model. Another solution is to build an equivariant neural network and thereby increasing the equivariance to a larger symmetry group. In this study, two convolutional neural networks and their respective equivariant counterparts are constructed and applied to the symmetry groups D4 and C8 to explore the impact on performance when removing and adding batch normalisation and data augmentation. The results suggest that data augmentation is irrelevant to an equivariant model and equivariance to more symmetries can slightly improve accuracy. The convolutional neural networks rely heavily on batch normalisation, whereas the equivariant models achieve high accuracy, although lower than with batch normalisation present.

In cancer research, there is a need to make accurate spatial measurements in multi-layered fluorescence microscopy images. Researchers would like to measure distances in and between biological objects such as nerves and tumours, to investigate questions which includes if nerve distribution in and around tumours can have a prognostic value in cancer diagnostics. This thesis is split into two parts, the first being: given arbitrary florescent images of cancer tissue samples, investigate the feasibility of automatically identifying nerves, tumours and blood vessels using classic image analysis. The second part is: given an image with identified objects, quantify their spatial data. By analysing 58 different cancer tissue samples we found that a modified Otsu method gives the most promising results for image segmentation. We found that non-verifiable objects and verifiable objects share the same pixel intensity distributions which implies that it is in general not possible to solely use thresholding methods to separate them from each other. For the spatial analysis, two measurement methods were introduced. An object based method that provides measurements from the edges of nerves to tumour edges, and a pixel based measurement method, which provides fraction based measurements that are comparable between different tissue samples.

Patients who are treated in an intensive care unit need continuous monitoring in orderfor clinicians to be prepared to intervene should a secondary event occur. For patientstreated at the neurointensive care unit (NICU) who have suffered a subarachnoid hemorrhage (SAH) this secondary event could be ischemia, resulting in a lack of blood flow.Blood flow can be measured using magnetic resonance imaging (MRI). The process is facilitated with a software called NOVA. Repeated measurements can therefore be performedas a way to monitor the patients, which in this context would be referred to as longitudinalmeasurements. As more data can be collected ways of analyzing and visualizing the datain a comprehensible way is needed. The aim of this thesis was therefore to develop and implement a method for analyzing and visualizing the longitudinal MR measurement data.With this aim in mind two research questions were relevant. The first one was how NOVAflow longitudinal measurements can be visualized to simplify interpretation by cliniciansand the second one was in what ways the longitudinal data can be analyzed. A graphicaluser interface (GUI) was created to present the developed analysis and visualization tool.Development of the tool progressed using feedback from supervisors and neurosurgeons.Visualization and analysis was done through plots of blood velocity and blood flow as themain component as well as a 2D vessel map. The final implementation showed multipleexamples of how the longitudinal data could be both visualized and analyzed. The resultstherefore provided a tool to analyze and visualize NOVA flow longitudinal measurementsin a way which was easily interpreted. Further improvements of the tool is possible andan area of improvement could involve increasing the adaptability of the tool.

Ultraljudsdiagnostik baseras på propagering av mekaniska vågor och används för att avbilda tvärsnitt av kroppen i realtid. Prestandan, med avseende på kontrast, upplösning och måttmätningar, av CONTEC CMS600B-3, en relativ billig maskin är av intresse. Hur volymen av en fantom, dess ytarea och frekvens på utskickade vågor påverkar uppvärmningen av vävnader är även av intresse. Det undersöktes med ultraljudsmaskinerna CONTEC CMS600B-3 och Philips Lumify för att få resultat som inte beror på endast en maskin.  Axiella upplösningen på CONTEC CMS600B-3 uppmättes med hjälp av ett gem till 0,61 mm och den laterala upplösningen till 1,27 mm med hjälp av ett snitt i cement. Maskinens måttmätningar hade en relativ avvikelse   beroende på mätning. Resultat för reflektionskoefficienten visade att ultraljudsmaskinen har en funktion som kompenserar för attenuering och förstärker signaler med låga amplituder.  Temperaturmätningarna undersöktes genom att skapa fantomer som efterliknar mänskliga vävnader med olika volymer och ytareor. En undersökning med ultraljudsmaskinerna visade att mer temperaturökning sker då ytarean ökas när volymen hålls konstant. Med avseende på säkerhet i temperaturökning, axiell upplösning och area/omkrets mått uppfyller CONTEC CMS600B-3 inte standarden och kan därmed inte användas inom sjukvården. / Diagnostics with ultrasound are based on propagation of mechanical waves and is used for imaging cross-section of the body in real-time. Performance, regarding contrast, resolution, and size measure, of CONTEC CMS600B-3, a relatively cheap machine is of interest. How volume of a phantom, its surface area, and frequency of the waves affects the heating of the tissues is also of interest. It was measured using ultrasound machines CONTEC CMS600B-3 and Philips Lumify to obtain results independent of the machine used.   The axial resolution of CONTEC CMS600B-3 was established with a paperclip to be 0.61 mm and the lateral resolution was measured to be 1.27 mm using concrete with a triangular slit. Measurements of the machine had a relative deviation was   depending on the measure. Results of the reflection coefficient indicated that CONTEC CMS600B-3 has a built-in function that compensates for loss of intensity due to attenuation and amplifies signals with lower amplitude to produce a B-mode image that the user can understand.  Temperature measurements were done on phantoms that mimic the human body with different volumes and surface areas. An investigation with ultrasound machines showed an increase in temperature with increased surface area as the volume is held constant. When looking at safety with temperature rise, axial resolution and area/circumference measurements, CONTEC CMS600B-3 does not meet the standard and therefore cannot be used in healthcare.

Segmentation of the knee cartilage is an important step for surgery planning and manufacturing patient-specific prostheses. What has been a promising technology in recent years is deep learning-based super-resolution methods that are composed of feed-forward models which have been successfully applied on natural and medical images. This thesis aims to test the feasibility to super-resolve thick slice 2D sequence acquisitions and acquire sufficient segmentation accuracy of the articular cartilage in the knee. The investigated approaches are single- and multi-contrast super-resolution, where the contrasts are either based on the 2D sequence, 3D sequence, or both. The deep learning models investigated are based on predicting the residual image between the high- and low-resolution image pairs, finding the hidden latent features connecting the image pairs, and approximating the end-to-end non-linear mapping between the low- and high-resolution image pairs. The results showed a slight improvement in segmentation accuracy with regards to the baseline bilinear interpolation for the single-contrast super-resolution, however, no notable improvements in segmentation accuracy were observed for the multi-contrast case. Although the multi-contrast approach did not result in any notable improvements, there are still unexplored areas not covered in this work that are promising and could potentially be covered as future work. / Segmentering av knäbrosket är ett viktigt steg för planering inför operationer och tillverkning av patientspecifika proteser. Idag segmenterar man knäbrosk med hjälp av MR-bilder tagna med en 3D-sekvens som både tidskrävande och rörelsekänsligt, vilket kan vara obehagligt för patienten. I samband med 3D-bildtagningar brukar även thick slice 2D-sekvenser tas för diagnostiska skäl, däremot är de inte anpassade för segmentering på grund av för tjocka skivor. På senare tid har djupinlärningsbaserade superupplösningsmetoder uppbyggda av så kallade feed-forwardmodeller visat sig vara väldigt framgångsrikt när det applicerats på verkliga- och medicinska bilder. Syftet med den här rapporten är att testa hur väl superupplösta thick slice 2D-sekvensbildtagningar fungerar för segmentering av ledbrosket i knät. De undersökta tillvägagångssätten är superupplösning av enkel- och flerkontrastbilder, där kontrasten är antingen baserade på 2D-sekvensen, 3D-sekvensen eller både och. Resultaten påvisar en liten förbättring av segmenteringnoggrannhet vid segmentering av enkelkontrastbilderna över baslinjen linjär interpolering. Däremot var det inte någon märkvärdig förbättring i superupplösning av flerkontrastbilderna. Även om superupplösning av flerkontrastmetoden inte gav någon märkbar förbättring segmenteringsresultaten så finns det fortfarande outforskade områden som inte tagits upp i det här arbetet som potentiellt skulle kunna utforskas i framtida arbeten.

Framgången av mönsterigenkänning inom AI (artificiell intelligens) har skapat höga förväntningar om att AI ska kunna appliceras inom vården, framför allt inom radiologi. Det danska företaget Radiobotics har utvecklat en maskininlärningsbaserad programvara som diagnostiserar knäartros, för att assistera vårdpersonalen i deras arbete. Denna AI-programvara vid namn RBknee analyserar en röntgenbild utifrån tre diagnostiska parametrar som förekommer vid knäartros, för att sedan sammanställa de radiologiska fynden i en skriftlig rapport tillsammans med en slutgiltig diagnos. För att få förståelse för hur RBknees analysförmåga påverkas av en bildkvalitetsförsämring undersöktes för vilken kontrast och brusnivå som RBknee genererar ett felaktigt utlåtande gällande de diagnostiska parametrarna och slutdiagnosen. Vidare undersöktes om graden av knäartros påverkade RBknee analysförmåga vid en bildkvalitetsförsämring. Ett bildunderlag med kliniskt tagna slätröntgenbilder av knän degraderades med avseende på kontrast och brus för att sedan analyseras av RBknee. Förändringar av RBknees utlåtande för de degraderade bilderna jämfört med originalbildens utlåtande sammanställdes och studerades. Resultatet visade att det inte gick att identifiera en specifik försämringsgrad av bildkvaliteten där RBknee genererade ett felaktigt utlåtande. RBknees förmåga att generera ett korrekt utlåtande var bättre vid en kontrastdegradering än vid en brusdegradering. Det konstaterades att en ökad brusnivå ökade risken för ett felaktigt utlåtande av RBknee, samt att brusets position på röntgenbilden hade en påverkan. Det gick även att fastställa att röntgenbilder av knän med en lägre grad av knäartros i högre grad riskerade att få felaktiga utlåtanden av RBknee. / The success of pattern recognition in AI (artificial intelligence) has brought high expectations for AI to be applied in healthcare, especially in radiology. A machine learning software for knee osteoarthritis diagnosis has been developed by the Danish company Radiobotics. The AI software, named RBknee, analyses digital radiographs and annotates osteoarthritis related findings. The findings, together with a conclusion, are compiled in a written report. RBknee is intended to assist healthcare professionals in radiographic analysis. How RBknees analytical ability is affected by a reduced image quality was studied by examining the contrast and noise level which cause RBknee to generate incorrect findings and conclusions. If the image quality reduction caused RBknees analytically ability to differ with different degrees of knee osteoarthritis, was also studied. The image quality of clinical digital radiographs of knees was reduced and analysed by RBknee. RBknees findings and conclusion were compared with the report of the original image, where the changes were compiled into tables. No specific reduction of image quality that restricted RBknee analytically ability was established in the study. An increased noise level seemed to increase the risk of receiving an incorrect report by RBknee. RBknees ability to generate correct report was better for contrast degraded images than for images with increased noise level. The position of the noise in the radiograph also seemed to have an impact on RBknees analytical ability. It was also possible to establish that knees with a lower degree of knee osteoarthritis were more likely to receive an incorrect report from RBknee.

Photon counting computed tomography scanners constitute a major improvement of the field of computed tomography, opening various prospective and enabling the decomposition of computed tomography images into different materials. The material decomposition algorithm, mapping photon counts to material pathlengths, relies on a forward model with Poisson statistics. This model though suffers from noise and residual bias due to its sensitivity to calibration errors and specificities in single-pixel responses that are not captured by the material decomposition model.           This study proposes a pixel-specific and projection-based correction of the residual bias in the material decomposition estimates using artificial neural networks trained for each pixel of the detector. The neural network models were trained under supervised learning using material decomposition calibration data, scans of PE and PVC slabs of various thicknesses acquired for the calibration of the model. This method aims at the mapping of the singularities of the pixels’ responses and correct them in the projection domain. The trained models were evaluated on a set of evaluation slabs and on scans of a water phantom, in order to assess performances of homogeneity and bias correction.           The implemented solution exhibited promising results for the correction of residual bias in single pixels without impairment of the noise levels. An array of trained neural networks demonstrates its ability to correct calibration and evaluation slab data while conserving pixel-to-pixel difference. The application of the correction to the water phantom however offered nuanced results which call for further investigation of the identified issues and induced improvements of the model.

4D-Flow MRI is a non-invasive imaging technique that can measure temporally resolved 3D images, capturing the flow/velocity in each pixel. The quality of the images and the temporal resolution largely depend on two factors. The acquisition protocol the MRI scanner uses and the reconstruction method used to go from signal to images. In MRI, the signal samples measured are the Fourier coefficients of the sought-after image, and reconstruction is an inverse problem, classically requiring sampling on at least Nyquist rate. Compressed sensing is a framework that allows for reconstruction from fewer samples than the Nyquist rate by incorporating other known information about the images. In this thesis, we evaluate the efficiency of Compressed Sensing for 4D-Flow MRI reconstruction for undersampled signals on synthetic data and compare it to classical reconstruction methods (Gridding and Viewshared Gridding). We specifically focus on the Locally Low Rank (LLR) regularization. The importance of initial-guess, or if it can be beneficial to estimate the temporal images by solving from the difference to the mean, is investigated. After calculating velocity profiles in vessels, we compare the reconstructed velocity profiles to the actual velocity profiles. We look at relative errors and pixel-wise maximum errors, as well as visual inspection. We introduce a velocity error metric aiming at capturing how accurate the reconstructed velocity profile is compared to our synthetic truth. We show that for good choices of regularization strength, the relative, maximum and velocity errors are significantly lower for the Compressed Sensing LLR method compared to the classical methods. We conclude that Compressed sensing with LLR regularization can significantly improve the reconstruction quality of 4D-Flow MRI data.

DeepPaint is a semi-automated segmentation tool that utilises a U-net architecture to performbinary segmentation. To maximise the model’s performance and minimise user time, it isadvisable to apply Transfer Learning (TL) and reuse a model trained on a similar segmentationtask. However, due to the sensitivity of medical data and the unique properties of certainsegmentation tasks, TL is not feasible for some applications. In such circumstances, SelfSupervised Learning (SSL) emerges as the most viable option to minimise the time spent inDeepPaint by a user. Various pretext tasks, exploring both corruption segmentation and corruption restoration, usingsuperpixels and square patches, were designed and evaluated. With a limited number ofiterations in both the pretext and downstream tasks, significant improvements across fourdifferent datasets were observed. The results reveal that SSL models, particularly those pretrained on corruption segmentation tasks where square patches were corrupted, consistentlyoutperformed models without pre-training, with regards to a cumulative Dice SimilarityCoefficient (DSC). To examine whether a model could learn relevant features from a pretext task, Centred KernelAlignment (CKA) was used to measure the similarity of feature spaces across a model's layersbefore and after fine-tuning on the downstream task. Surprisingly, no significant positivecorrelation between downstream DSC and CKA was observed in the encoder, likely due to thelimited fine-tuning allowed. Furthermore, it was examined whether pre-training on the entiredataset, as opposed to only the training subset, yielded different downstream results. Asexpected, significantly higher DSC in the downstream task is more likely if the model hadaccess to all data during the pretext task. The differences in downstream segmentationperformance between models that accessed different data subsets during pre-training variedacross datasets.