Hypothesis Unilateral congenital conductive hearing impairment in ear canal atresia leads to atrophy of the gray matter of the contralateral primary auditory cortex or changes in asymmetry pattern if left untreated in childhood. Background Unilateral ear canal atresia with associated severe conductive hearing loss results in deteriorated sound localization and difficulties in understanding of speech in a noisy environment. Cortical atrophy in the Heschl's gyrus has been reported in acquired sensorineural hearing loss but has not been studied in unilateral conductive hearing loss. Methods We obtained T1w and T2w FLAIR MRI data from 17 subjects with unilateral congenital ear canal atresia and 17 matched controls. Gray matter volume and thickness were measured in the Heschl's gyrus using Freesurfer. Results In unilateral congenital ear canal atresia, Heschl's gyrus exhibited cortical thickness asymmetry (right thicker than left, corrected p = 0.0012, mean difference 0.25 mm), while controls had symmetric findings. Gray matter volume and total thickness did not differ from controls with normal hearing. Conclusion We observed cortical thickness asymmetry in congenital unilateral ear canal atresia but no evidence of contralateral cortex atrophy. Further research is needed to understand the implications of this asymmetry on central auditory processing deficits.

Tractography has become an indispensable part of brain connectivity studies. However, it is currently facing problems with reliability. In particular, a substantial amount of nerve fiber reconstructions (streamlines) in tractograms produced by state-of-the-art tractography methods are anatomically implausible. To address this problem, tractogram filtering methods have been developed to remove faulty connections in a postprocessing step. This study takes a closer look at one such method, Spherical-deconvolution Informed Filtering of Tractograms (SIFT), which uses a global optimization approach to improve the agreement between the remaining streamlines after filtering and the underlying diffusion magnetic resonance imaging data. SIFT is not suitable for judging the compliance of individual streamlines with the acquired data since its results depend on the size and composition of the surrounding tractogram. To tackle this problem, we propose applying SIFT to randomly selected tractogram subsets in order to retrieve multiple assessments for each streamline. This approach makes it possible to identify streamlines with very consistent filtering results, which were used as pseudo-ground truths for training classifiers. The trained classifier is able to distinguish the obtained groups of complying and non-complying streamlines with the acquired data with an accuracy above 80%.

Incorporating neuroimaging-revealed structural details into finite element (FE) head models opens vast new opportunities to better understand brain injury mechanisms. Recently, growing efforts have been made to integrate fiber orientation from diffusion tensor imaging (DTI) into FE models to predict white matter (WM) tract-related deformation that is biomechanically characterized by tract-related strains. Commonly used approaches often downsample the spatially enriched fiber orientation to match the FE resolution with one orientation per element (i.e., element-wise orientation implementation). However, the validity of such downsampling operation and corresponding influences on the computed tract-related strains remain elusive. To address this, the current study proposed a new approach to integrate voxel-wise fiber orientation from one DTI atlas (isotropic resolution of 1 mm(3)) into FE models by embedding orientations from multiple voxels within one element (i.e., voxel-wise orientation implementation). By setting the responses revealed by the newly proposed voxel-wise orientation implementation as the reference, we evaluated the reliability of two previous downsampling approaches by examining the downsampled fiber orientation and the computationally predicted tract-related strains secondary to one concussive impact. Two FE models with varying element sizes (i.e., 6.4 +/- 1.6 mm and 1.3 +/- 0.6 mm, respectively) were incorporated. The results showed that, for the model with a large voxelmesh resolution mismatch, the downsampled element-wise fiber orientation, with respect to its voxel-wise counterpart, exhibited an absolute deviation over 30 across the WM/gray matter interface and the pons regions. Accordingly, this orientation deviation compromised the computation of tract-related strains with normalized root-mean-square errors up to 30% and underestimated the peak tract-related strains up to 10%. For the other FE model with finer meshes, the downsampling-induced effects were lower, both on the fiber orientation and tract-related strains. Taken together, the voxel-wise orientation implementation is recommended in future studies as it leverages the DTI-delineated fiber orientation to a larger extent than the element-wise orientation implementation. Thus, this study yields novel insights on integrating neuroimaging-revealed fiber orientation into FE models and may better inform the computation of WM tract-related deformation.

Tractography aims at describing the most likely neural fiber paths in white matter. A general issue of current tractography methods is their large false-positive rate. An approach to deal with this problem is tractogram filtering in which anatomically implausible streamlines are discarded as a post-processing step after tractography. In this chapter, we review the main approaches and methods from the literature that are relevant for the application of tractogram filtering. Moreover, we give a perspective on the central challenges for the development of new methods, including modern machine learning techniques, in this field in the next few years.

The opportunity to non-invasively probe the structure and function of different parts of the human body makes medical imaging an indispensable tool in clinical diagnostics and related fields of research. Especially neuroscientists rely on modalities like structural or functional Magnetic Resonance Imaging, Computed Tomography or Positron Emission Tomography to study the human brain in vivo. But also in clinical routine, diagnosis, screening or follow-up of different pathological conditions build upon the use of neuroimaging.

Computational solutions are essential for the analysis of medical images. While in the case of conventional photography the recorded signal comprises the actual image, most medical imaging devices require the reconstruction of an image from the acquired data. However, not only the image formation, but also further processing tasks to assist doctors or researchers in the interpretation of the data and eventually in subsequent decision making, rely more and more on automation. Typical tasks range from locating and measuring objects in a single patient, e.g. a particular organ, a tumour or a specific region in the brain, to comparing such measurements over time between groups consisting of large numbers of subjects. Automated solutions for these scenarios are required to model complex relations of data in the presence of acquisition noise and subject variability while assuring a tractable computational demand.

Traditionally, the development of computational algorithms for medical imaging problems focused on rule-based strategies. Explicitly defined rules that encode the knowledge of the developer are characteristic for such approaches. Within the last decade, this paradigm began to change and learning-based models dramatically gained in popularity. These rely on fitting a complex model to large amounts of data samples, often annotated, which are representative for a particular problem. Instead of manually designing the sought-after solution, it is ‘learned from the data’. While these models have shown enormous potential, they also pose important questions for method developers. How can I get hold of enough data? How much data is enough? How can I obtain proper annotations?

This thesis comprises six studies covering the development and the application of methods along the whole pipeline of medical image analysis. Studies I and II propose different extensions to the method of tensor voting to make it applicable in specific medical imaging problems. Studies III–V address the use of modern machine learning techniques, in particular neural networks, in the field of tractography. Notably, the challenge of obtaining adequately annotated data samples is a topic in Study V. In Study VI, a prospective neuroimaging study of unilateral ear canal atresia in adults is presented, covering the application of methods from data acquisition to group comparison. Overall, the compiled works contributed, in one way or the other, to the non-invasive extraction of knowledge from the human body through automated processing of medical images.

Möjligheten att från utsidan undersöka struktur och funktion hos olika delar av människokroppen har gjort medicinsk avbildning till ett oumbärligt verktyg i klinisk diagnostik och relaterade forskningsområden. Särskilt inom neurovetenskap är forskarna starkt beroende av metoder som strukturell eller funktionell magnetresonanstomografi (MRT), datortomografi (DT) eller positronemissionstomografi (PET) för att studera hjärnan hos den levande människan. Men också i rutinsjukvården bygger diagnostik, hälsokontroll eller uppföljning av olika sjukdomstillstånd på avbildning av hjärnan.

Beräkningsmetoder är oundgängliga för att analysera medicinska bilder. I motsats till konventionell fotografi, där den insamlade signalen innehåller själva bilden, kräver de flesta medicinska avbildningsmetoder att bilden rekonstrueras från insamlade data. Men det är inte bara bildalstringen, utan även den fortsatta behandlingen för att stödja det efterföljande beslutsfattandet, som är mer och mer automatiserade. Typiska uppgifter kan handla om att lokalisera och mäta strukturer i den enskilda patienten – t.ex. ett visst organ, en tumör eller en del av hjärnan – eller att jämföra sådana mätningar över tid mellan grupper bestående av ett stort antal personer. Automatiserade lösningar för dessa uppgifter krävs för att modellera komplexa relationer mellan data som är behäftade med insamlingsbrus och individuell variation utan att beräkningarna blir ohanterligt krävande.

Av tradition har utvecklingen av beräkningslösningar för medicinska avbildningsproblem fokuserat på regelbaserade strategier, ett arbetssätt som kännetecknas av explicit definierade regler som omsätter utvecklarens kunskaper. Under det senaste årtiondet har detta paradigm börjat ändras, och inlärningsbaserade modeller har ökat dramatiskt i popularitet. Dessa bygger på att en komplex modell anpassas till stora datamängder, ofta försedda med något slags anteckningar av erfarna praktiker (annoteringar), som kännetecknar ett specifikt problem. I stället för att manuellt konstruera den eftersökta lösningen, blir den inlärd från data. Samtidigt som dessa modeller har en enorm potential, ställer de utvecklarna inför viktiga frågor: Hur ska jag få tag i tillräckligt mycket data? Hur mycket data är tillräckligt? Hur ska jag få annoteringar av hög kvalitet?

Denna avhandling omfattar sex studier som täcker utveckling och tillämpning av metoder genom hela den medicinska bildbehandlingskedjan. Studie I och II föreslår olika utvidgningar av metoden tensorröstning för att göra den tillämpbar på specifika medicinska avbildningsproblem. Studie III–V behandlar användningen av moderna maskinlärningstekniker, mer specifikt neuronnät, inom området traktografi (avbildning av nervbanor i hjärnan). Utmaningen i att erhålla tillräckliga mängder annoterade data är centralt i Studie V. Studie VI utgör en prospektiv hjärnavbildningsstudie på vuxna med outvecklad inre hörselgång och innefattar tillämpning av metoder från datainsamlingtill gruppjämförelser. Sammantaget har alla de ingående arbetena, på ett eller annat sätt, bidragit till icke-invasiv generering av kunskap om människokroppen genom automatiserad medicinsk bildbehandling.

Supervised machine learning (ML) algorithms have recently been proposed as an alternative to traditional tractography methods in order to address some of their weaknesses. They can be path-based and local-model-free, and easily incorporate anatomical priors to make contextual and non-local decisions that should help the tracking process. ML-based techniques have thus shown promising reconstructions of larger spatial extent of existing white matter bundles, promising reconstructions of less false positives, and promising robustness to known position and shape biases of current tractography techniques. But as of today, none of these ML-based methods have shown conclusive performances or have been adopted as a de facto solution to tractography. One reason for this might be the lack of well-defined and extensive frameworks to train, evaluate, and compare these methods. In this paper, we describe several datasets and evaluation tools that contain useful features for ML algorithms, along with the various methods proposed in the recent years. We then discuss the strategies that are used to evaluate and compare those methods, as well as their shortcomings. Finally, we describe the particular needs of ML tractography methods and discuss tangible solutions for future works.

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.

This paper focuses on predicting a single step of streamline tractography from diffusion magnetic resonance imaging data by using different predictors based on neural networks. We train 18 different classifiers in order to assess the effect of including neighbourhood information in the learning step or as a post processing step. Moreover, the performance using four different post processing approaches as well as the variation of the number of classes resulting in a total of 60 experimental configurations are assessed. Further, a comparison to 12 regression-based networks is performed and the effect of including several streamline steps in the network input is investigated. All networks are trained and tested on the ISMRM 2015 tractography challenge data. Our results do not indicate a clear improvement when using neighbouring data (regardless if it used as an input or as a post processing). Also, the linear interpolation of the diffusion data does not outperform the less expensive nearest neighbour approach. As opposed to that, using a linear model on top of the output of the classifiers is beneficial and—in combination with at least 200 classes—resulted in a similar performance as the regression approach. Finally, providing the networks with additional curvature information led to a clear improvement of prediction performance. Our analysis of accuracy based on average angular errors suggests that also considering spatial location in the learning step might further improve machine learning-based streamline tractography algorithms.