The future of artificial intelligence (AI) in medical imaging is full of promise as well as challenges. Computer vision can support clinicians and researchers in the diagnosis of diseases and in the pursuit of new routes of treatment. With the current advances in computer vision approaches and the availability of a greater variety of data, you may wonder what medical imaging applications will benefit from artificial intelligence? How can we apply artificial intelligence in medical imaging to improve diagnosis or treatment?
In August 2018, the National Institutes of Health (NIH) organized a workshop in Bethesda, Maryland, to investigate new methods of employing AI in the analysis of medical images. One goal was to develop a roadmap for the creation of new medical image analysis tools, prioritizing certain unexploited avenues of research.
AI engineers and researchers report rapid progress in the creation of computer vision applications and achieved performances that are comparable to that of humans or that even outperform humans in analyzing medical images. Sophisticated techniques are used for tasks such as computer-aided object detection, radiogenomics, computer-aided classification, segmentation, triage, noise reduction, and image reconstruction.
The teams of researchers participating in the NIH workshop defined research priorities that will advance the field of AI applications in medical imaging. These research priorities include the creation of new machine learning methods, new methods for anonymizing data, new automated image labeling and annotation methods, as well as new image reconstruction methods.
Novel machine learning approaches can help in the creation of models that are pre-trained for medical imaging data. Moreover, the researchers hope to create machine learning models that can potentially explain their decisions to a user. Validation methods for image de-identification and anonymization will facilitate secure and easy ways to share clinical imaging datasets. Automated image labeling and annotation methods can assist in the creation of imaging reports, while image reconstruction methods can be used to create human interpretable images from raw data.
The workshop report explained how the innovations that were road mapped will help to increase the number of publicly available, validated datasets that in turn can be used to create novel algorithms and AI models. In order for medical image datasets to be of use, a pre-processing of the imaging datasets may be required to quickly label/annotate image data. AI algorithms can speed up the annotation of image data in the form of autonomous or semi-autonomous labeling procedures. Pre-trained model architectures that are designed for medical imaging datasets have the potential to dramatically reduce the amount of data that needs to be exchanged between institutions.
By laying out the primary research pathways in the field of AI applications in medical imaging, the researchers have paved the way for further technological advances that will facilitate diagnoses and will thus improve overall health. Let’s take a look at some examples of how artificial intelligence can be leveraged in a medical imaging context.
AI approaches for the improvement of medical images range from image distortion corrections, to tissue revealing functions, and thus facilitate clinical decision making.
Researchers from the ETH Zurich and the University of Zurich have employed machine learning methods for the optimization of optoacoustic imaging. Optoacoustic imaging is a medical imaging technique that has only been introduced recently and that operates similarly to ultrasound. Optoacoustic imaging uses sounds waves and is for example applied in the diagnosis of breast cancer, in the categorization of skin lesions, in the visualization of blood vessels, and in the monitoring of brain activity. One limitation of the optoacoustic technique is the image quality that heavily depends on the number of sensors that are used in the image acquisition. Machine learning algorithms can be used to reduce the number of sensors that is needed for the acquisition of high-quality images. This in turn reduces the device costs while improving diagnostic accuracy and/or the image acquisition speed.
The Zurich research team was able to use artificial intelligence to correct distortions that stem from the optoacoustic approach. For this purpose, the research team captured high-quality images using 512 sensors, and analyzed these high-quality images using a neural network. The research team then created images using only 128, respectively 32 sensors. Acquiring images with less sensors created distortions in the images. The network that was trained on the high-quality images was then used to repair and enhance the distorted images that were captured using the limit number of sensors.
The image quality of an optoacoustic image improves in direct proportion to the number of sensors that is used in the image acquisition.However, the image quality also improves for specific sensor arrangements: Here the sensors need to be well distributed around the object to be represented, and thus need to cover various viewpoints and overall wide field of view. However, the proposed machine learning algorithm is able to improve the quality of images that were captured using a narrow field of view as well.
In addition, the machine learning algorithms proposed by the Zurich team can be used to capture images representing functional and molecular information without the need for contrast agents. Contrast agents are chemical compounds that are used in combination with medical imaging modalities such as X-rays and MRIs to track changes in metabolic functions or in the blood flow. Using the AI enhanced optoacoustic method, chemical contrast agents are no longer needed for these visualizations. This facilitates the monitoring of symptoms and diseases such as the monitoring of tissue oxygenation, which is used as a cancer landmark. The AI enhanced optoacoustic method can also be used to measure the lipid content in blood vessels, which can be used for an early detection of cardiovascular diseases.
Even the most sophisticated artificial intelligence approaches are rendered useless if the data is not correctly prepared for an analysis by the underlying neural network. Image annotation facilitates the training of such a neural network. If annotations contain sloppy mistakes or if they are incorrect, these errors will negatively affect the performance of the image classifier. For this reason, you should consider to outsource your image annotation tasks to professionals who will ensure a high quality data as the optimal foundation for your algorithms.