Recent advancements in machine learning (ML) have led to a dramatic increase in AI capabilities for medical diagnostic tasks. Despite technical advances, developers of predictive AI models struggle to integrate their work into routine clinical workflows. Inefficient human-AI interactions, poor sociotechnical fit and a lack of interactive strategies for dealing with the imperfect nature of predictions are known factors contributing to this lack of adoption.

User-centred design methods are typically aimed at discovering and realising desirable qualities in use, pragmatically oriented around finding solutions despite the limitations of material- and human resources. However, existing methods often rely on designers possessing knowledge of suitable interactive metaphors and idioms, as well as skills in evaluating ideas through low-fidelity prototyping and rapid iteration methods—all of which are challenged by the data-driven nature of machine learning and the unpredictable outputs from AI models.

Using a constructive design research approach, my work explores how we might design systems with AI components that aid clinical decision-making in a human-centred and iterative fashion. Findings are derived from experiments and experiences from four exploratory projects conducted in collaboration with professional physicians, all aiming to probe this design space by producing novel interactive systems for or with ML components.

Contributions include identifying practical and theoretical design challenges, suggesting novel interaction strategies for human-AI collaboration, framing ML competence for designers and presenting empirical descriptions of conducted design processes. Specifically, this compilation thesis contains three works that address effective human-machine teaching and two works that address the challenge of designing interactions that afford successful decision-making despite the uncertainty and imperfections inherent in machine predictions.

Finally, two works directly address design-researchers working with ML, arguing for a systematic approach to increase the repertoire available for theoretical annotation and understanding of the properties of ML as a designerly material.

Recent technical advances has made machine learning (ML) a promising component to include in end user facing systems. However, user experience (UX) practitioners face challenges in relating ML to existing user-centered design processes and how to navigate the possibilities and constraints of this design space. Drawing on our own experience, we characterize designing within this space as navigating trade-offs between data gathering, model development and designing valuable interactions for a given model performance. We suggest that the theoretical description of how machine learning performance scales with training data can guide designers in these trade-offs as well as having implications for prototyping. We exemplify the learning curve's usage by arguing that a useful pattern is to design an initial system in a bootstrap phase that aims to exploit the training effect of data collected at increasing orders of magnitude.

Whole slide pathology images from regional lymph node metastasis in colon adenocarcinoma produced at Region Gävleborg Clinical Pathology and Cytology department. Consists of fifty chronologically consecutive cases.

Background: Recent advancements in machine learning (ML) bring great possibilities for the development of tools to assist with diagnostic tasks within histopathology. However, these approaches typically require a large amount of ground truth training data in the form of image annotations made by human experts. As such annotation work is a very time-consuming task, there is a great need for tools that can assist in this process, saving time while not sacrificing annotation quality. Methods: In an iterative design process, we developed TissueWand – an interactive tool designed for efficient annotation of gigapixel-sized histopathological images, not being constrained to a predefined annotation task. Results: Several findings regarding appropriate interaction concepts were made, where a key design component was semi-automation based on rapid interaction feedback in a local region. In a user study, the resulting tool was shown to cause substantial speed-up compared to manual work while maintaining quality. Conclusions: The TissueWand tool shows promise to replace manual methods for early stages of dataset curation where no task-specific ML model yet exists to aid the effort.

What if the trust in the output of a predictive model could be acted upon in richer ways than a simple binary decision of accept or reject? Designing assistive AI tools for medical specialists entails supporting a complex but safety-critical decision process. It is common that decisions in this domain can be decomposed to a combination of many smaller decisions. In this paper, we present Verification Staircase – a design strategy that can be used for such scenarios. The verification staircase is when multiple interactive assistive tools are combined to allow for a nuanced amount of automation to aid the user. This can support a wide range of prediction quality scenarios, spanning from unproblematic minor mistakes to misleading major failures. By presenting the information in a hierarchical way, the user is able to learn how underlying predictions are connected to overall case predictions, and over time, calibrate their trust so that they can choose the appropriate level of automatic support.

Objective: Digital pathology is today a widely used technology, and the digitalization of microscopic slides into whole slide images (WSIs) allows the use of machine learning algorithms as a tool in the diagnostic process. In recent years, “deep learning” algorithms for image analysis have been applied to digital pathology with great success. The training of these algorithms requires a large volume of high-quality images and image annotations. These large image collections are a potent source of information, and to use and share the information, standardization of the content through a consistent terminology is essential. The aim of this project was to develop a pilot dataset of exhaustive annotated WSI of normal and abnormal human tissue and link the annotations to appropriate ontological information. 

Materials and Methods: Several biomedical ontologies and controlled vocabularies were investigated with the aim of selecting the most suitable ontology for this project. The selection criteria required an ontology that covered anatomical locations, histological subcompartments, histopathologic diagnoses, histopathologic terms, and generic terms such as normal, abnormal, and artifact. WSIs of normal and abnormal tissue from 50 colon resections and 69 skin excisions, diagnosed 2015-2016 at the Department of Clinical Pathology in Linköping, were randomly collected. These images were manually and exhaustively annotated at the level of major subcompartments, including normal or abnormal findings and artifacts. 

Results: Systemized nomenclature of medicine clinical terms (SNOMED CT) was chosen, and the annotations were linked to its codes and terms. Two hundred WSI were collected and annotated, resulting in 17,497 annotations, covering a total area of 302.19 cm², equivalent to 107,7 gigapixels. Ninety-five unique SNOMED CT codes were used. The time taken to annotate a WSI varied from 45 s to over 360 min, a total time of approximately 360 h. 

Conclusion: This work resulted in a dataset of 200 exhaustive annotated WSIs of normal and abnormal tissue from the colon and skin, and it has informed plans to build a comprehensive library of annotated WSIs. SNOMED CT was found to be the best ontology for annotation labeling. This project also demonstrates the need for future development of annotation tools in order to make the annotation process more efficient.

The dataset consists of 101 H&E-stained colon whole slide images (WSI) - 52 abnormal and 49 benign cases. All significant abnormal findings identified are outlined and categorized into 15 types such as hyperplastic polyp, high grade adenocarcinoma and necrosis. Other tissue components such as mucosa, submucosa, as well as the surgical margin are delineated to create a complete histological map. In total, 756 separate annotations have been made to segment the different tissue structures and link them to ontological information.

Whole slide pathology images from regional lymph node metastasis in colon adenocarcinoma produced at Region Gävleborg Clinical Pathology and Cytology department and Region Östergötland Clinical Pathology department. Annotations for AI training produced as part of AIDA clinical fellowship project investigating AI decision support in metastasis detection.