Jennie Karlsson will defend her PhD-thesis: "Trustworthy Classification of Medical Images - Deep Learning for Breast Cancer and Coronary Artery Disease Applications"

Opponent: 

Prof. Ida-Maria Sintorn, Uppsala University

Grading committee:

Elisabeth Wetzer, UiT

Joakim Lindblad, Uppsala University

Jennifer Alvén, Chalmers

Abstract:

Imaging plays a central role in modern diagnostic workflows and early detection of diseases is often strongly associated with improved patient outcomes. As artificial intelligence (AI) becomes more integrated into society, new opportunities arise in several fields, including medical imaging. In this context, deep learning, a branch of AI, has emerged as a powerful approach for automatic image analysis. Such methods could offer automatic disease assessment, with the potential to improve diagnostic accuracy, reduce workload, expand access to healthcare services, and facilitate earlier detection. This thesis investigates deep learning for automatic disease assessment, using breast cancer and coronary artery disease (CAD) as clinical applications to study disease classification and model trustworthiness.
Breast cancer has a profound impact on women worldwide, with survival rates being significantly lower in low- and middle-income countries compared to high-income countries. A major contributing factor to this disparity is the lack of timely diagnosis. In this thesis, point-of-care ultrasound (POCUS), a portable and relatively inexpensive imaging technique, is investigated in combination with deep learning-based classification models as a support tool for breast cancer assessment. Similarly, CAD poses a substantial global health burden. The evaluation of CAD consists of several steps, starting with non-invasive procedures and commonly followed by more invasive interventions when disease is suspected. This thesis investigates the use of deep learning to assess the disease from non-invasive imaging with the potential to identify patients at risk requiring further assessment, and to prevent unnecessary procedures.
The implementation of deep learning models for medical image classification presents several challenges, including ensuring accurate disease prediction, addressing data scarcity, and establishing model trustworthiness. In this thesis convolutional neural networks (CNNs) are implemented for disease assessment, yielding promising predictive results. To address data scarcity, several strategies are explored. These include conventional data augmentation to increase data diversity, synthetic image generation using cycle-consistent adversarial networks (CycleGANs) for cross-domain image translation, and autoencoders for feature learning from partially annotated datasets. The methods proved to be useful, with data augmentation showing particular importance.
Furthermore, several approaches to increase the trustworthiness of a model are investigated, including out-of-distribution (OOD) detection, uncertainty quantification (UQ), and explainability using saliency maps. OOD detection is used to identify unfamiliar samples that could compromise a model’s predictions, while UQ is used to estimate prediction uncertainty. Saliency maps are used to improve explainability by providing a visual interpretation of the model’s prediction. Additionally, this thesis includes a validation study comparing a deep learning model with radiologists for assessment of breast POCUS images. This is an important step toward developing robust and clinically applicable deep learning-based diagnostic support tools