Deep learning for medical data

This thesis revisits the fundamental components of deep learning and evaluates their application in the context of medical image analysis. It identifies three main challenges where deep learning falls short in this domain: the integration of expert knowledge, the leveraging of unlabeled data, and the estimation of predictive uncertainty. The thesis is structured into parts that address these challenges respectively.
In part 1, the thesis introduces a novel deep learning model that incorporates expert knowledge through roto-reflective equivariance to improve the accuracy and robustness of medical imaging tasks, specifically in the detection of metastatic tissue in histopathology slides. The proposed model outperforms traditional CNN architectures and demonstrates robustness to input perturbations. Next follows an exploration on how to motivate the deep learning community to focus on real-world medical problems by presenting PCam, a dataset derived from the Camelyon16 challenge. PCam is structured to resemble common deep learning benchmarks and demonstrates that improvements on this dataset translate to improvements on the larger Camelyon16 benchmark.
Part 2 explores the benefits of self-supervised representation learning through Contrastive Predictive Coding (CPC) and proposes Contrastive Perturbative Predictive Coding (C2PC) that enhances CPC's performance by incorporating specific medical imaging augmentations.
Part 3 of the thesis addresses the challenge of estimating predictive uncertainty, crucial for high-risk medical decision-making. It introduces a novel variational inference method that leverages multinomial distributions over quantized latent variables. The proposed method exhibits competitive performance in uncertainty estimation and risk assessment compared to existing methods.
The thesis concludes that by addressing the identified challenges, deep learning can be better suited for medical imaging tasks. It demonstrates that expert knowledge can be effectively integrated into deep learning models, that leveraging unlabeled data through self-supervised learning can enhance model performance, and that predictive uncertainty can be improved with more flexible variational inference methods.