Deep Learning for Improved Diagnosis of Pathologies in Wireless Capsule Endoscopy with Focus on Data Efficiency and Transparency
Doctoral thesis
Permanent lenke
https://hdl.handle.net/11250/3100287Utgivelsesdato
2023Metadata
Vis full innførselSamlinger
Sammendrag
Over the past decade the use of deep learning in different application areas has increased steadily. These areas include automated decisions in industries to autonomous vehicles, defence, and even medical. Each field gives rise to its own unique challenges and considerations that are required to be studied and addressed so as to best cater to its development for the good of the society. This thesis is dedicated to understanding and advancing these aspects for the emerging diagnostic modality of Wireless Capsule Endoscopy (WCE) in the medical domain. It explores the great potential for deep learning based algorithms to aid and assist medical experts in carrying out diagnoses for patients in a time and cost effective manner. The complexity of medical information contained within WCE images presents several challenges that this thesis works through systematically. The techniques span three key objectives: 1. advance computer aided diagnosis in the face of data and label adversity, 2. generate synthetic data and evaluate its quality with a view to discovering factors that makeup WCE data; and 3. explore the automation of explanation generation for diagnosis.
Firstly, we tackle the label and data adversity problem by proposing two methodologies for pathology classification. These methods rely on leveraging domain-specific knowledge to act as guidance in situations where supervision through labels is not available. Here a multi-pathology classification benchmark has been proposed for WCE. Secondly, we seek to understand if disease biomarkers can be learned and visually simulated such that synthetic data that appears realistic to experts can be generated from completely unlabeled data. This contributes to new knowledge about what biomarkers exist in the data and how they can be utilized for simulating plausible disease variations in the future. The outcome is an atlas of synthetically generated WCE images that span a number of pathological and non-pathological variations. Lastly, we integrate this knowledge to gain insights into the relationship between these variations and the semantic concepts that medical practitioners are familiar with, in order to develop more readily comprehensible explanations in the context of WCE.
The potential impact of techniques presented in this thesis extends beyond the realm of WCE and into other domains. Proposed methodologies address some of the key challenges faced in diagnosing WCE pathologies, including the scarcity of labeled data and the difficulty of interpreting complex images, both of which could have significant implications for improving diagnostic accuracy and ultimately patient outcomes.