Machine Learning in Cardiothoracic Radiology: From Medical Data Curation to Clinical Application

The thesis is organized into 8 chapters. Its main theme is the creation and investigation of Machine Learning (ML) algorithms for a broad spectrum of tasks in the field of cardiothoracic radiology.

This includes data curation, a fundamental step at the beginning of most scientific projects, exemplified by an algorithm for radiology report classification (Chapter 2): a CNN trained on 2801 impression sections of CT pulmonary angiogram reports was able to classify radiology reports with an accuracy of 99.1% according to whether they describe the presence of pulmonary embolism or not.

Furthermore, image recognition algorithms for detection (Chapters 3 & 4) and segmentation (Chapter 5) of cardiothoracic findings in cross-sectional imaging are presented: A 3D CNN with ResNet architecture was trained to detect pulmonary embolism on CT pulmonary angiograms, using 28000 examinations, and reached a sensitivity of 92.7% and specificity of 95.5% on an independent test set. This algorithm is currently used in multiple hospitals around the world. Another CNN’s sensitivity for the detection of rib fractures in whole-body trauma CTs was 97.4% on a per examination level (specificity: 91.5%). An algorithm pipeline with two CNN algorithms at its core (faster RCNN and ResNet) was able to detect lung cancer of the T1 category with a sensitivity of 90.4% and showed excellent segmentation performance.

Finally, a model using cardiothoracic imaging biomarkers automatically extracted from chest CT with machine learning to predict the clinical course of patients infected with SARS-CoV-2 is discussed (Chapter 6). It proved to be an excellent classifier for differentiation of patients eventually needing ICU-care vs. those who did not, with an area-under-the-curve of 0.88, using solely information available at the time of initial presentation at the hospital.

Thereby, the chapters of this thesis follow a chronologic sequence of tasks, beginning with data curation to finding detection to finding segmentation to application of extracted information for clinical decision support. Finally, Chapter 7 summarizes and discusses methodological insights gained during the abovementioned ML projects and provides a framework of best practices for creation and evaluation of machine learning algorithms in clinical practice.

The thesis is framed by an introduction (Chapter 1), which includes a brief introduction to ML, discusses the relevance of ML to radiology, and provides examples of current ML applications in radiology with a focus on cardiothoracic imaging. And by a concluding chapter (Chapter 8) that summarizes the main findings and discusses future directions of Machine Learning in radiology with a focus on cardiothoracic imaging.