Lukas Folle
Data-Driven Derivation of Rheumatological Disease Characteristics from Medical Images
Rheumatologists have various tools at hand that allow characterizing the diseases of their patients. A multitude of those tools requires human annotation, which in some cases is time-intensive and error-prone. This poses a unique challenge, that can be addressed by modern pattern recognition advancements. In this work, a brief overview of rheumatic diseases was provided. Thereafter, imaging methods commonly used in rheumatology were described. Lastly, the fundamentals of deep learning and neural networks were explained, which build the foundation of the methods applied during the works, that are part of this thesis. As a first step to reducing time and increasing consistency, the workflow to estimate the density of patient bones using computed tomography (CT) was addressed. Tracking the bone density of patients over time provides necessary information for clinicians to adequately adapt patient therapy. As this requires manual annotation of the bone in CT images, it is time-intensive and error-prone. By using a segmentation network to assist humans in this step, the clinical workflow could be considerably accelerated while still generating high-quality results. The second step of this thesis regarded the classification of patients based on magnetic resonance images (MRI) and CT scans to different groups of diseases. As rheumatic diseases specifically target the soft tissue of patients and the bone structures, those imaging techniques are of great utility in the rheumatological practice. However, disease mechanisms are still not fully understood. Thus, in this thesis, the ability to classify patients into the different disease entities of rheumatic diseases using the information present only in MRI and CT was investigated. Based on both imaging techniques, with the help of neural networks, classifiers with solid performance were trained. Interestingly, certain features in the shape of the bones, that have been described in previous works, influenced the predictions of the network to a high degree. In the last work of this thesis, the classification and prediction of the severity of nail psoriasis, a condition that occurs during severe disease stages of psoriasis, was studied. Currently, the grading of nail psoriasis is mostly based on the nail psoriasis severity index (NAPSI). However, even though this can help clinicians in their decision process, it is not clinically used due to the high time-intensiveness of the grading process. By acquiring a dataset of hand photos and the corresponding labels, a neural network was trained to predict the NAPSI automatically. By providing access to the complete system online, the applicability of the approach for other clinicians was greatly improved and enables for the first time the usage of the NAPSI not just in clinical studies but also in clinical practice. At the end of this thesis, the limitations of the proposed works were discussed. A common aspect across all works was the necessity for a greater amount of training data. Specifically, the disease entity classification works reached a solid performance, but would likely benefit greatly from an increasing amount of training data. The nail psoriasis classification work reached very good performance, but future work, that would increase utility for patients would be the development of a mobile application that runs the proposed method and tracks the estimated disease severity over time. Briefly, rheumatic diseases and their representation in CT, MRI, and hand photos were investigated in this thesis. With the help of neural networks, time-consuming workflows could be considerably accelerated and new insights into the disease mechanics were gained.