Artificial Intelligence (AI) has been revolutionizing the healthcare industry by automating tasks, improving predictions, and being able to analyze large amounts of data. One of the biggest applications has been in terms of analyzing medical diagnoses in order to detect patterns and predict potential health problems early on. However, there’s a challenge that comes to this: being “black-box models”, they often lack transparency in their results, and this raises concerns regarding understanding the model’s decision-making processes. To address this, Explainable Artificial Intelligence (XAI) was created. It can offer clear and understandable explanations for AI actions, through visual explanations. This dissertation focuses on utilising XAI methods to enhance transparency, fairness, and safety in a medical image classification, more specifically in a customised model for detecting pulmonary diseases. The first phase of the project involves a systematic literature review to analyse existing research on XAI in healthcare applications. The review focuses on articles published between 2016 and 2022, obtained from various databases. The selection criteria includes relevance to XAI in healthcare, and the final eligible articles underwent thorough scrutiny and statistical analysis. From this systematic search three different XAI methods were chosen to be applied in our model. In the implementation phase, GRAD-CAM, LIME, and RISE were integrated into two model approaches: a pre-trained and fully-trained model for chest X-ray image classification. Two of the XAI techniques provided visual explanations for individual predictions, improving interpretability. The dissertation is concluded by evaluating the performance of both the model and the XAI techniques. The pre-trained model achieved high performance in detecting pulmonary diseases. Two of the XAI techniques had similar results, enhancing the trust in the model patterns and decision-making process. This model aims to improve patient care and outcomes. The ongoing journey towards developing trustworthy and interpretable AI models in healthcare must be kept as a priority when enhancing healthcare delivery.

A necessidade de se ter, em certos contextos clínicos, um doente hipocoagulado em ambulatório, esteve durante muitos anos centrada na utilização dos antagonistas da vitamina K, merecendo destaque nesse âmbito a varfarina. Os antagonistas da vitamina K, também designados por anticoagulantes clássicos, apesar de constituírem um padrão na anticoagulação oral, apresentam múltiplas limitações que obrigam a uma rigorosa e frequente monitorização da terapêutica. Estes fármacos apresentam ainda um elevado risco de efeitos adversos e de interações farmacológicas. Estas desvantagens da terapêutica anticoagulante clássica, levou que novos anticoagulantes orais fossem desenvolvidos e testados. Estes fármacos têm efetivamente apresentado um perfil de eficácia e segurança superior àquele que se verifica com a utilização dos referidos antagonistas da vitamina K, com múltiplas vantagens do ponto de vista clínico e do doente.