The impressive technical advances seen for machine learning algorithms in combination with the digitalization of medical images in the radiology and pathology departments show great promise in introducing powerful image analysis tools for image diagnostics. In particular, deep learning, a subfield within machine learning, has shown great success, advancing fields such as image classification and detection. However, these types of algorithms are only used to a very small extent in clinical practice. 

One reason is that the unique nature of radiology and pathology images and the clinical setting in which they are acquired poses challenges not seen in other image domains. Differences relate to capturing methods, as well as the image contents. In addition, these datasets are not only unique on a per-image basis but as a collective dataset. Characteristics such as size, class balance, and availability of annotated labels make creating robust and generalizable deep learning methods a challenge. 

This thesis investigates how deep learning models can be trained for applications in this domain, with particular focus on histopathology data. We investigate how domain shift between different scanners causes performance drop, and present ways of mitigating this. We also present a method to detect when domain shift occurs between different datasets. Another hurdle is the shortage of labeled data for medical applications, and this thesis looks at two different approaches to solving this problem. The first approach investigates how labeled data from one organ and cancer type can boost cancer classification in another organ where labeled data is scarce. The second approach looks at a specific type of unsupervised learning method, self-supervised learning, where the model is trained on unlabeled data. For both of these approaches, we present strategies to handle low-data regimes that may greatly increase the availability to build deep learning models for a wider range of applications. 

Furthermore, deep learning technology enables us to go beyond traditional medical domains, and combine the data from both radiology and pathology. This thesis presents a method for improved cancer characterization on contrast-enhanced CT by incorporating corresponding pathology data during training. The method shows the potential of im-proving future healthcare by intergraded diagnostics made possible by machine-learning technology. 

Unsupervised learning has made substantial progress over the last few years, especially by means of contrastive self-supervised learning. The dominating dataset for benchmarking self-supervised learning has been ImageNet, for which recent methods are approaching the performance achieved by fully supervised training. The ImageNet dataset is however largely object-centric, and it is not clear yet what potential those methods have on widely different datasets and tasks that are not object-centric, such as in digital pathology.While self-supervised learning has started to be explored within this area with encouraging results, there is reason to look closer at how this setting differs from natural images and ImageNet. In this paper we make an in-depth analysis of contrastive learning for histopathology, pin-pointing how the contrastive objective will behave differently due to the characteristics of histopathology data. Using SimCLR and H&E stained images as a representative setting for contrastive self-supervised learning in histopathology, we bring forward a number of considerations, such as view generation for the contrastive objectiveand hyper-parameter tuning. In a large battery of experiments, we analyze how the downstream performance in tissue classification will be affected by these considerations. The results point to how contrastive learning can reduce the annotation effort within digital pathology, but that the specific dataset characteristics need to be considered. To take full advantage of the contrastive learning objective, different calibrations of view generation and hyper-parameters are required. Our results pave the way for realizing the full potential of self-supervised learning for histopathology applications. Code and trained models are available at https://github.com/k-stacke/ssl-pathology.

Clinical deployment of systems based on deep neural networks is hampered by sensitivity to domain shift, caused by e.g. new scanners or rare events, factors usually overcome by human supervision. We suggest a correct-then-predict approach, where the user labels a few samples of the new data for each slide, which is used to update the network. This few-shot meta-learning method is based on Model-Agnostic Meta-Learning (MAML), with the goal of training to adapt quickly to new tasks. Here we adapt and apply the method to the histopathological setting by identifying a task as a whole-slide image with its corresponding classification problem. We evaluated the method on three datasets, while purposefully leaving out-of-distribution data out from the training data, such as whole-slide images from other centers, scanners or with different tumor classes. Our results show that MAML outperforms conventionally trained baseline networks on all our datasets in average accuracy per slide. Furthermore, MAML is useful as a robustness mechanism to out-of-distribution data. The model becomes less sensitive to differences between whole-slide images and is viable for clinical implementation when used with the correct-then-predict workflow. This offers a reduced need for data annotation when training networks, and a reduced risk of performance loss when domain shift data occurs after deployment.

The high capacity of neural networks allows fitting models to data with high precision, but makes generalization to unseen data a challenge. If a domain shift exists, i.e. differences in image statistics between training and test data, care needs to be taken to ensure reliable deployment in real-world scenarios. In digital pathology, domain shift can be manifested in differences between whole-slide images, introduced by for example differences in acquisition pipeline - between medical centers or over time. In order to harness the great potential presented by deep learning in histopathology, and ensure consistent model behavior, we need a deeper understanding of domain shift and its consequences, such that a model's predictions on new data can be trusted. This work focuses on the internal representation learned by trained convolutional neural networks, and shows how this can be used to formulate a novel measure - the representation shift - for quantifying the magnitude of model specific domain shift. We perform a study on domain shift in tumor classification of hematoxylin and eosin stained images, by considering different datasets, models, and techniques for preparing data in order to reduce the domain shift. The results show how the proposed measure has a high correlation with drop in performance when testing a model across a large number of different types of domain shifts, and how it improves on existing techniques for measuring data shift and uncertainty. The proposed measure can reveal how sensitive a model is to domain variations, and can be used to detect new data that a model will have problems generalizing to. We see techniques for measuring, understanding and overcoming the domain shift as a crucial step towards reliable use of deep learning in the future clinical pathology applications.

Domain shift is a significant problem in histopathology. There can be large differences in data characteristics of whole-slide images between medical centers and scanners, making generalization of deep learning to unseen data difficult. To gain a better understanding of the problem, we present a study on convolutional neural networks trained for tumor classification of H&E stained whole-slide images. We analyze how augmentation and normalization strategies affect performance and learned representations, and what features a trained model respond to. Most centrally, we present a novel measure for evaluating the distance between domains in the context of the learned representation of a particular model. This measure can reveal how sensitive a model is to domain variations, and can be used to detect new data that a model will have problems generalizing to. The results show how learning is heavily influenced by the preparation of training data, and that the latent representation used to do classification is sensitive to changes in data distribution, especially when training without augmentation or normalization.