Scientists Use Machine Learning Methods to Improve Optoacoustic Imaging

Machine learning techniques have now been used by researchers at the University of Zurich and at ETH Zurich to enhance optoacoustic imaging.

This comparatively new medical imaging method can be employed for various applications, like the diagnosis of breast cancer, characterization of skin lesions, examining brain activity, and visualizing blood vessels.

Conversely, the quality of the rendered images depends largely on the distribution and number of sensors utilized by the device — that is, more numbers of sensors will translate to better image quality.

With the help of the latest technique created by the ETH team, the number of sensors can be considerably reduced without compromising on the resulting quality of the images. This means, device cost can be reduced, imaging speed can be increased, or diagnosis can be enhanced.

Optoacoustics is analogous to ultrasound imaging in certain respects. In ultrasound imaging, ultrasonic waves sent into the body by a probe are reflected by the tissue. The returning sound waves are detected by sensors integrated into the probe, and a picture of the interiors of the body is eventually created.

In the case of optoacoustic imaging, extremely short laser pulses are delivered to the tissue, where they get absorbed and changed into ultrasonic waves. The waves — similar to ultrasound imaging — are identified and changed into images.

Correcting for Image Distortions

Headed by Daniel Razansky, Professor of Biomedical Imaging at ETH Zurich as well as the University of Zurich, the research group looked for a method to improve the image quality of economical optoacoustic devices that contain only a few ultrasonic sensors.

To accomplish this feat, the researchers initially utilized a self-developed sophisticated optoacoustic scanner integrated with 512 sensors that provided images of excellent quality. They used an artificial neural network to examine these pictures. This network managed to learn the traits of high-quality images.

Afterward, the team discarded most of the sensors, leaving just 32 or 128 sensors. This had a detrimental impact on the quality of the images. Since data was not available, distortions called streak-type artifacts were seen in the images. However, it turned out that the earlier trained neural network was able to fix these distortions considerably, thus bringing the quality of the images closer to the measurements achieved with all the 512 sensors.

The image quality in optoacoustics increases with the number of sensors utilized and also when the data is recorded from as many directions as possible — that is, image quality is directly proportional to the size of the sector in which the sensors are organized around the object. In addition, the developed machine learning algorithm was also effective in enhancing the quality of images recorded from only a narrowly circumscribed sector.

This is particularly important for clinical applications, as the laser pulses cannot penetrate the entire human body, hence the imaged region is normally only accessible from one direction.

Daniel Razansky, Professor of Biomedical Imaging, ETH Zurich

Facilitating Clinical Decision Making

However, the researchers emphasized that their technique is not restricted to optoacoustic imaging. The technique does not work on the raw recorded data but does only on the reconstructed images. Therefore, it can also be used in other imaging methods.

You can basically use the same methodology to produce high-quality images from any sort of sparse data.

Daniel Razansky, Professor of Biomedical Imaging, ETH Zurich

He further explained that physicians are usually faced with the challenge of studying inferior quality images obtained from patients.

“We show that such images can be improved with AI methods, making it easier to attain more accurate diagnosis,” added Razansky.

For Razansky, the study serves as an excellent example of what current techniques of artificial intelligence can be utilized for.

Many people think that AI could replace human intelligence. This is probably exaggerated, at least for the currently available AI technology. It can’t replace human creativity, yet may release us from some laborious, repetitive tasks.

Daniel Razansky, Professor of Biomedical Imaging, ETH Zurich

In their present study, the researchers utilized an optoacoustic tomography device that was tailor-made for small animals, and then used images from mice to train the machine learning algorithms.

Razansky added that the team’s next step will be to apply the technique to optoacoustic images from human patients.

Revealing Tissue Function

Several imaging methods like MRI, X-ray, or ultrasound are different from other optoacoustics (also called photoacoustics). They are mainly suited for visualizing the body’s anatomical changes.

In order to get more functional data, for example, regarding metabolic changes or blood flow, radioactive tracers or contrast agents have to be injected into the patient prior to imaging. On the other hand, the optoacoustic technique can visualize molecular and functional data without having to administer any contrast agents.

One such example is local variations that occur in tissue oxygenation — a major landmark of cancer that can be employed for early diagnosis. Another potential disease marker is lipid content in blood vessels, which can help in the early diagnosis of cardiovascular diseases.

However, it must be noted that since the light waves utilized in optoacoustic imaging — in contrast to other waves — do not completely enter the human body, the technique can only be used for examining tissues to a depth of few centimeters under the skin.