Abstract & Overview
Objectives and outline of the thesisTo address the mentioned challenges, several deep learning models for cardiac related tasks were developed in this thesis.In several projects a cooperation between technical and medical experts from renowned scientific and medical institutes was formed to process and analyze data from unprecedentedly large clinical cohorts.Chapter 2 presents a DL system for automatic cardiovascular risk prediction.The proposed system consists of four consecutive steps (1) to localize and (2) segment the heart in a CT scan, (3) segment coronary calcium and (4) calculate a risk score.The system was developed using cardiac ECG-gated CT scans from the community based, observational Framingham Heart Study (FHS) 55 .It was tested in over 20,000 CT scans from 4 distinct clinical cohorts, including low-dose chest screening CTs of asymptomatic individuals from the National Lung Screening Trial (NLST) 56 as well as cardiac gated CTs of individuals with acute chest pain from the Prospective Multicenter Imaging Study for Evaluation of Chest Pain (PROMISE) 57 and Rule Out Myocardial Infarction using Computer Assisted Tomography II study (ROMICAT-II) 58 .The scans were gathered by over 200 participating medical sites and manual segmentations were provided by experienced medical experts from the Cardiovascular Imaging Research Center at the Massachusetts General Hospital.Accuracy, compared to the gold standard of expert human readers, was assessed in 5,521 subjects across all four cohorts.The results demonstrate that deep learning methods can automate cardiovascular risk prediction from medical images acquired in several clinical scenarios.These observations provide a rationale to implement this technique in both screening and hospital settings to improve population health, at high speed and low costs.Chapter 3 presents a reliable and accurate deep learning system for fully automatic heart segmentation in non-contrast enhanced cardiac ECG-gated CT and low-dose chest screening CT scans.The model was trained on high quality scans from FHS 59,60 , with manual segmentations provided by experienced medical experts.The system was tested on two large and distinct clinical cohorts including cardiac ECG-gated CTs from PROMISE 57 as well as low-dose chest screening CTs from NLST 56 .In the test cohorts totalling 1,534 CT scans acquired in 226 medical sites, performance, generalizability, and applicability of the proposed deep learning system was assessed.The system is suitable for a vast array of research and medical applications.By making the code open-source and providing the fully trained model to the public without restrictions, this study has the potential to accelerate clinical research and improve medical treatment.Chapter 4 presents a study to determine the association of whole heart volume with major adverse cardiovascular events, adjusting for traditional measures of cardiovascular risk and coronary artery disease (CAD) characteristics on CT.Furthermore, a subgroup analysis across CAD categories was performed to determine whether whole heart volume had discriminatory capacity incremental to atherosclerosis cardiovascular disease risk score and CT-derived CAD characteristics. General introduction and outline of the thesis 13 Chapter 1Chapter 5 investigates the relationship between epicardial adipose tissue volume and traditional cardiovascular risk factors, CT-derived coronary artery disease characteristics, and incident adverse events in symptomatic individuals with increased cardiovascular risk from PROMISE 57 .Chapter 6 investigates whether a deep learning system developed in cardiovascular radiology can be used for radiation oncology treatment planning.Therefore, a deep learning system for whole heart segmentation was developed using multi-center data including dedicated cardiac CTs and low dose chest screening CTs, with manual segmentations from expert cardiovascular radiologists.This system was validated in an independent real-world dataset including 5,677 breast cancer patients which were treated at the Dana-Farber and Brigham and Women's Cancer Center between 2008-2018.The performance of the deep learning system was compared to radiation oncology experts as well as to heart segmentations used in the clinic for treatment delivery.This study may serve as proof of principle to repurpose and leverage AI applications for optimizing patient care and reducing healthcare disparities across specialties, institutions, and countries.Chapter 7 presents a summary of this work and next steps, future improvements and applications.