1. To understand the basics of machine learning.
2. To appreciate the opportunities to improve quality in radiology.
3. To redefine the professional role of the radiologist.
Novel developments in computer science and performance have prepared the ground for the implementation of so-called artificial intelligence algorithms and have computers learn “deeply” from the analysis of big data. This session will provide an introduction into machine based learning algorithms touching reflecting opportunities, but also limitations. Obviously, quantitative image analysis will be a promising field for machine learning and the integration of the “machines” in the workflow of radiology will lead to major challenges to generate benefits for patients, the health care system and their economics as well as our profession. Done properly, machines will help us to cope with the ever increasing volume of imaging exams and scans, increase quality and safety, while radiologists will generate the value from imaging and the added value provided by the machines to strengthen our role as physicians.
1. To learn about the principles of machine learning.
2. To appreciate the current performance of deep learning in radiology and beyond.
3. To learn about the limitations due to the complexity of radiology.
The application of deep learning to medical image analysis has resulted in much-improved performance for many algorithms. There is early evidence that for many specific, narrowly defined detection and interpretation tasks, computers perform on par with or even better than radiologists. A requirement is that these deep learning algorithms are trained on a sufficiently large amount of high-quality annotated data. Larger multi-center validation studies are still needed to demonstrate that these algorithms work correctly on the variety of real-world data, obtained with different scanners and protocols. In this talk, I will showcase several examples of high-performing deep learning algorithms and argue that the radiological community should play a more active role in (1) identifying which tasks should be automated; (2) evaluating and validating automated systems in a fair and comparative manner; (3) collecting the high-quality annotated training and evaluation data; (4) clearly defining the role of human readers together with automated systems.
1. To become familiar with first clinical applications.
2. To consolidate knowledge about the integration in the clinical workflow.
3. To appreciate the future roadmap and its impact on training of young radiologists.
The recent advances in artificial intelligence, particularly deep learning algorithm, have shown performances superior to those of human in image recognition of non-medical fields. Accordingly, many researchers and companies have tried to apply artificial intelligence in radiology. Artificial intelligence can be applied in various aspects of medical imaging such as noise reduction, artefact reduction, lesion segmentation, detection, classification, and differential diagnosis. Early studies have shown promising results in several applications such as detecting nodule or tuberculosis on chest radiograph. Furthermore, this technique can applied in finding similar cases from the database, predicting disease outcome and so on. In this talk, general concept of artificial intelligence and early results of various applications of artificial intelligence in medial image will be introduced.
1. To become familiar with the opportunities of big data analysis.
2. To understand that quality of big data is pivotal.
3. To learn how to generate "evidence" from big data analysis in radiology.
Deep learning is currently an extremely active research area and has the potential to be the biggest game changer in medical imaging since the advent of digital imaging. It requires massive amount of training dataset with accurate annotation; however, annotation of medical images is particular difficult and resource consuming as compared to other examples outside of medicine. Potentially induced error and bias due to imperfect annotation and/or inconsistency increase further the needed sample size. Several sources of medical images can be considered. While clinically acquired data are broadly available, they are often not interlinked with other data and face the issue of a high variability. A potentially more suitable source for big data could be large, population-based cohort studies such as the Rotterdam Study, the UK BioBank or the German National Cohort, since they have a high level of data quality and consistency regarding imaging, and they further frequently linked with clinical outcomes and annotations. However, deep learning in medical imaging remains a great challenge, not only because of the necessity for big data.