Artificial intelligence-based framework for characterizing atherosclerotic plaques in coronary arteries using intravascular optical coherence tomography: Application to calcified lesions.

Atherosclerosis is a chronic inflammatory disease of coronary arteries, affecting the structure and function of their wall. It is characterized by the progressive accumulation of lipids within the arterial wall, triggering a continuous inflammatory response. This process gradually leads to reduced arterial wall elasticity and lumen narrowing. Before clinical effects become visible, atherosclerotic plaques develop slowly and insidiously over decades (30-50 years). Investigating the plaque’s characteristics throughout its progression may provide valuable insights for preventing adverse events associated with long-term atherosclerosis. Intravascular optical coherence tomography (OCT) has revolutionised the understanding of atherosclerosis in coronary arteries, allowing early identification of vulnerable plaques. OCT is a catheter-based modality that evaluates the cross-sectional and 3-dimensional microstructure of blood vessels at exceptional resolution (12-40 microns). Despite its benefits, the use of OCT is hampered by the complexity of image interpretation, which requires specialised training and time. This factor is a limitation to its widespread use in daily clinical practice and can lead to delays in patient care. Artificial intelligence (AI) has the potential to address some of these challenges, enabling the identification of elements that may be unnoticed by the human eye or through traditional analysis, significantly automating image processing and accelerating the extraction of quantitative data. In this thesis work, AI-based methods were employed to develop a framework for identifying atherosclerotic plaque features in OCT coronary artery images. To achieve this goal, a deep learning-based pipeline has been implemented, consisting in two steps: (1) an image pre-processing step aiming at image cleaning, based on a convolutional autoencoder, and (2) the identification of the desired plaque feature by a convolutional neural network (CNN) architecture. This pipeline has been specifically applied to the identification of calcium since is a good marker of atherosclerosis. The methodology used to build the above-mentioned pipeline, first involved the use of a convolutional autoencoder, a type of neural network suitable for structured data such as images. Its application was aimed to pre-process the images to be used as input for the calcium classifier by removing parts that were not informative on the presence of calcium (catheter, blood and other artefacts). The following step was the application of a CNN classifier based on transfer learning, a powerful technique allowing to use gained knowledge from pre-trained models to solve new classification problems more efficiently and accurately. The performance of the model was evaluated using metrics such as precision, recall, F1 score, and area under the ROC curve (AUC), as the available dataset was highly unbalanced, with the calcium class being the minority class. Overall, the model showed fair performance and allowed the identification of calcium within the arterial wall. Further improvements of the current model include increasing its effectiveness and precision. In conclusion, this study proposes a versatile workflow that is, in principle, suitable for the identification of different plaque components. Furthermore, this approach opens the door to future studies identifying processes such as plaque rupture, which is crucial for preventing acute cardiovascular events such as stroke, embolism or myocardial infarction.