The goal of Stream 1 is to develop physics- and artificial intelligence-based methods to reduce and simplify scan times, increase the number of quantifiable parameters, and improve disease characterization, ensuring reproducibility across different equipment and medical centers. We aim to overcome current limitations in medical imaging—such as the lack of comprehensive quantification, integration with other clinical data, and high operational costs—through the development of innovative techniques that integrate AI, imaging physics, and multiparametric data, from acquisition to clinical diagnosis.
Research
Research Areas
Research Line
Stream 1: Intelligent fully quantitative medical imaging
Research Line
Stream 2: AI-enabled physics-informed and synergistic medical imaging
The goal of Stream 2 is to develop physics-informed AI methods that leverage synergies across medical imaging modalities (ultrasound, magnetic resonance, etc.) throughout the entire process (from acquisition to diagnosis and prognosis), and between different specifications (conventional and lower-cost medical imaging equipment) to improve efficiency (faster scans/lower radiation dose), effectiveness (capturing a broader spectrum of disease-related processes), and accessibility (solutions for both conventional and lower-cost systems).
Research Line
Stream 3: Explainable AI for medical imaging and non-imaging patient data integrated analysis, reporting and diagnosis
The goal of Stream 3 is to develop self-supervised Explainable Artificial Intelligence (XAI) methods that integrate medical imaging with other patient data (physiological sensors, clinical reports, risk factors, etc.) to provide accurate, reliable, and interpretable AI-assisted clinical analysis. Developing methods that are more explainable for the end-user would significantly support decision-making and build trust among healthcare professionals.
Research Line
Stream 4: Population-based phenotyping for cardiovascular, liver disease and cancer
The goal of Stream 4 is to improve the diagnosis and prognosis of cardiovascular, hepatic, and oncological diseases by translating the technical advances from Streams 1 to 3, as well as by correlating AI-derived imaging phenotypes with clinical descriptors, non-imaging data, and outcomes. We also aim to reduce bias in AI model research due to underrepresentation in available datasets by creating a Chilean medical imaging database (iHEALTH-BB).
Research Projects
Research project
Fully Quantitative Magnetic Resonance Imaging
Medical imaging requires sequential acquisition of various (mostly qualitative) anatomical/functional images and subjective expert interpretation to correlate imaging findings with a diagnosis. Quantitative medical imaging is a major research focus in magnetic resonance imaging. However, current solutions are still limited by their accuracy, the restricted number of quantifiable parameters and by lack of reproducibility and standardization. Moreover, image quality can be affected by system imperfections, motion and other confounding factors, hindering analysis and interpretation. In this project we aim to develop accurate and easy to use/interpret (“one-click” exam) MRI able to provide comprehensive quantitative tissue characterization (e.g. fibrosis, inflammation) from a single and fast scan.
Research project
Multiparametric MR for the assessment of acute cardiac toxicity induced by thoracic radiotherapy
Today, cancer ranks as one of the leading causes of death. Despite the large number of novel available therapies, radiotherapy remains as the most effective non-surgical method to cure cancer patients. In fact, approximately 50% of all cancer patients receive some type of radiotherapy and among these, 60% receive radiotherapy-treatment with a curative intent. However, as occurs with any other oncological therapy, radiotherapy treated patients may experience toxicity side effects that range from moderate to severe. Among these, cardiotoxicity represents a significant threat for premature death. Current methods evaluate cardiotoxic damage based on volumetric changes in the Left Ventricle Ejection Fraction (LVEF). Indeed, a 10% drop in LVEF is commonly used as an indicator of cardiotoxicity. More recently, several novel techniques have been developed that significantly improve specificity and sensitivity of heart’s volumetric changes and early detection of cardiotoxicity even in asymptomatic patients. In this project we will use novel multi-parametric MRI to characterize the myocardial tissue and strain analysis to develop early biomarkers of cardiotoxicity in a cohort of patients that will receive radiotherapy.
Research project
Automatic Report Generation from Medical Images - Database and Methods
The task of Medical Image Report Generation can be of great support for physicians. From a computational perspective, the Medical Image Report Generation task can be described as follows: given as input one or more medical images of a patient, a text report is output that is as similar as possible to one generated by a radiologist. From a machine learning point of view, creating a system that performs such a task would require learning a generative model from instances of reports written by radiologists. This project aims at creating the technology to automatically generate reports from medical images.
