Preprocessing Approaches for Informed Machine Learning in Medical Imaging : Emerging Trends in Digital Health, Summer Term 2023

Background: In recent years, the integration of machine learning algorithms and neural networks in medical image classification has become increasingly significant. These technologies offer promising capabilities for assisting healthcare professionals in interpreting various types of medical images, such as X-rays, MRIs, CT scans, and pathology slides. However, the effectiveness of these algorithms heavily relies on the availability of sufficient and high-quality training data, which often presents a bottleneck due to limited and inaccessible patient datasets.
Objective: This paper aims to explore how informed machine learning approaches can improve the accuracy and effectiveness of medical image classification models by
integrating domain knowledge into the imaging algorithm. Specifically, the research investigates preprocessing steps utilized in this process, focusing on both the input data and the domain knowledge.
Methods: A systematic review of 80 publications was conducted to identify and analyze the preprocessing approaches employed in integrating domain knowledge into medical image classification tasks. The study categorized the preprocessing steps into three main categories: "Data Preprocessing," "Knowledge Preprocessing," and "Knowledge-Enhanced Data Preprocessing." Additionally, the characteristics of the domain knowledge were examined to understand the relationships between preprocessing steps and the structure of available knowledge.
... mehr
Results: The analysis revealed that the adaptation of existing image preprocessing pipelines and machine learning image classification tasks required minimal changes. Instead, the emphasis was on preprocessing the domain knowledge and designing a compatible architecture. The findings highlighted the importance of effectively incorporating specialized knowledge and expert input to enhance model accuracy and effectiveness.
Conclusion: Informed machine learning approaches, which integrate domain knowledge into medical image classification algorithms, offer promising opportunities to
improve diagnostic accuracy and efficiency in healthcare. By focusing on preprocessing steps related to both data and domain knowledge, researchers can design more effective and compatible models, ultimately advancing the capabilities of medical image analysis systems.