Machine learning techniques, especially artificial neural networks, share many ideas and features with classical (that is, non-ML) methods for solving inverse problems. Examples are underlying Tikhonov-type optimization problems and the interpretation of deep neural networks as iterative methods structured like typical forward-backward splitting. In the talk we discuss those similarities and draw conclusions on possible directions for future research. Chances and limitations of ML techniques are discussed in the context of inverse problems from medical imaging. Of particular interest will be susceptibility weighted MR imaging (SWI).

The accurate detection and localization of natural or artificial structures in medical images is essential for effective diagnostics and surgical planning. In particular, determining the pose of artificial markers in MRI images is a foundational step for subsequent spatial adjustments, such as the registration between imaging modalities and with surgical devices. Manual detection and localization of these markers can be tedious and time-consuming, which has prompted the exploration of reliable, and highly automated approaches that can significantly reduce the need for human interaction. In recent years, automatic approaches based on neural networks have shown remarkable success in the detection and semantic segmentation of natural, anatomical structures. In contrast to these structures, the geometry of artificial markers is typically known, which enables the development of relatively simple algorithms that can perform well without the need for complex neural network architectures. The challenge often lies in the incomplete and inhomogeneous representation of the markers within the MRI images, due to noise, distortions, artifacts and further image defects. In this talk, we will explore different automatic approaches to MRI marker localization from a practical perspective, including conventional image processing pipelines utilizing basic methods such as convolution filters or connected component analysis and labeling as well as approaches based on neural networks. By addressing the benefits and challenges of using these methods, we gain a better understanding of their potential applications and impact on clinical image processing workflows.

In recent years the availability of multi-omics data, had a great impact on medical research. Such high-dimensional data-sets contain molecular as well as radiological variables from genomics, epigenomics, transcriptomics, proteomics, metabolomics, microbiomics and radiomics. The challenge when working with this kind of data is to find a subset of meaningful variables in order to deduce disease specific characteristics or make predictions on clinical endpoint variables. The process of eliminating non-informative and redundant features is called feature selection. Feature selection can be considered a necessary pre-processing for the actual modeling step in which statistical or machine-learning models are built in order to conduct classification tasks or time-to-event-data analysis. Given the pre-selected subset of features and a variety of candidate models, finding the most accurate as well as informative model is thereafter the remaining challenge also referred to as model selection. The goal of model selection is eliciting a parsimonious model, which uses only a small set of explanatory variables, which can then be considered as clinical covariates or biomarkers and in turn provide information how the treatment of the disease can be improved. Further, model selection is a step to prevent misleading conclusions due to possible over-fitting of the data inherent noise. In this talk, we present recent methodologies in feature and model selection. An introduction of rather simple feature selection techniques like statistical filter and classifier performance focused methods is followed by a description of more sophisticated regularization and shrinkage methods as well as the utilization of decision tree analysis algorithms. Finally we discuss how statistical as well as machine-learning models can benefit from the application of various information criteria for model selection.

The presentation focuses on a critical analysis of the performance of artificial neural networks (ANNs) in the context of semantic segmentation of organs as reported in scientific reports. In recent years, extensive research has been performed on combining loss functions and developing non-trainable layers for ANNs in order to optimize the boundary regions of semantic segmentation. These boundaries are particularly crucial for various segmentation tasks such as the detection of water retention in the lungs for COVID-19 diagnosis, the localization of organs at risk in radiotherapy treatment planning or the identification of white matter hyperintensities in the brain. With the U-Net, Ronneberger et al. have designed a powerful network architecture for any type of segmentation task. With a reported IOU (intersection over union) of $92\%$ and $77.5\%$ for the datasets used in the original work, respectively, it was far superior to the follow-up network. On this basis, U-Net was used for many segmentation tasks in the following years. In the field of semantic organ segmentation, the U-Net and U-Net-like artificial neural networks achieved very high accuracies. Since then, the reported accuracy has hardly improved. However, the underlying calculation of accuracy is misleading, as the improvements in recent years have been aimed at improving boundary regions, but these are usually unfavourably proportionate to the inner area.Hence, improvements in boundary segmentation accuracy has only a marginal impact on the overall accuracy. To illustrate this, we will look at the reported performance and improvements of artificial neural networks in both single-task and multi-task applications, using lung segmentation as an example. Typical evaluation methods, specifically Dice coefficient or Hausdorff distance, will be presented with current values and improvements. An overview of new evaluation methods and a discussion of the current way of reporting will also be addressed.