Candidate:
Maria Helena Sampaio de Mendonça Montenegro e Almeida
Date Time and Location:
16 June 2026, 14:30, Sala de Atos, Faculdade de Engenharia da Universidade do Porto
President of the Jury:
Carlos Miguel Ferraz Baquero-Moreno (PhD), Full Professor at the Department of Informatics Engineering of the Faculdade de Engenharia da Universidade do Porto
Members:
Peter Johannes Schüffler (PhD), Professor at the Technical University of Munich, Germany;
Carlos Jorge Andrade Mariz Santiago (PhD), Research Assistant at the Robotics and Engineering Systems Laboratory (LARSyS) and Invited Assistant Professor of the Department of Electrical and Computer Engineering of the Instituto Superior Técnico da Universidade de Lisboa;
Jaime dos Santos Cardoso (PhD), Full Professor of the Department of Electrical and Computer Engineering of the Faculdade de Engenharia da Universidade do Porto (Supervisor);
Luís Filipe Pinto de Almeida Teixeira (PhD), Associate Professor of the Department of Informatics Engineering of the Faculdade de Engenharia da Universidade do Porto.
Abstract:
Artificial Intelligence models have been extensively applied to medical image analysis tasks over the past years, achieving outstanding results. However, the obscure reasoning of the models and the lack of supportive evidence causes both clinicians and patients to distrust their predictions, hindering their adoption in clinical practice. In recent years, the research community has focused on developing explanations capable of revealing a model’s reasoning. Among the various types of explanations, case-based explanations emerged as particularly intuitive for medical practitioners. While these types of explanations have been widely researched, they still possess limitations that compromise their real-world application. The main goal of this thesis is to overcome the limitations of medical case-based explanations, enabling their deployment in clinical practice.
To identify the main weaknesses of case-based explanations, we conduct a literature review on models that provide such explanations in healthcare. Through the analysis of existing works, we verify that the explanations raise privacy concerns by sharing sensitive images of patients, and lack the interactiveness required for clinicians to engage with the explanation. Furthermore, most works do not evaluate nor clinically validate the explanations through user studies with clinicians. Targeting these limitations, we propose deep generative models to obtain privacy-preserving, controllable and interactive explanations to explain the decisions of deep learning models.
To propose a privacy-preserving system for safely sharing case-based explanations, we start by reviewing the literature on image anonymisation techniques and identifying their vulnerabilities. In particular, we identify vulnerabilities by proposing two privacy attacks: a membership inference attack targeting deep generative models that generate synthetic images, and a superimposed image decomposition model to reverse the common anonymisation strategy of image averaging. After outlining the requirements that a privacy-preserving system for medical case-based explanations must fulfil, we propose privacy-preserving models capable of simultaneously anonymising medical images and generating counterfactual explanations. The proposed models rely on disentangled representation learning to separate identity and clinical traits, enabling their individual manipulation. Through this strategy, we can also separate and manipulate causal factors related with the clinical task to obtain controllable counterfactual explanations. To obtain interactive counterfactual explanations, we propose models to manipulate regions of medical images according to a segmentation mask, measuring their impact on a prediction. To evaluate the proposed models, we perform experiments on medical datasets like chest radiographs, verifying the models’ capacity to manipulate medical images for the purposes of generating explanations.
Finally, to show the wide applicability of the proposed models beyond their original purpose of generating case-based explanations, we apply them as decision-support systems in healthcare. More specifically, we adapt the models for predicting the aesthetic outcomes of breast cancer treatment, to facilitate the patients’ choice of treatment. Furthermore, we conduct a clinical study with breast surgeons from various health institutes to assess the predictions of the proposed models.
To conclude, through various contributions on explainable Artificial Intelligence and decision-support systems, this thesis contributes towards the safe use of trustworthy and privacy-preserving Artificial Intelligence models in healthcare.
