"Artificial Intelligence Will Make Medicine Better in the Long Run" – Biophotonics.World

Image source: Leibniz IPHT

By: Sven Dring

Leibniz IPHT is increasingly focusing on artificial intelligence and learning systems. Thomas Bocklitz is heading the new research department "Photonic Data Science". We asked him how AI could help shape the future of diagnostics

What new possibilities does Photonic Data Science open up for diagnostics?

Photonic Data Science is a potpourri combining mathematical and statistical methods with algorithms and domain knowledge to translate measurement data into useful information. We usually translate photonic data into biomedical for example diagnostic information. By translating with the computer, robust diagnostic information can be extracted. Tiny details in complex data can be made useful for diagnostics. This opens up new possibilities for diagnostics.

Artificial intelligence (AI) then helps to evaluate this data. Which technologies researched at the institute are based on AI?

In the laser-based rapid test of infectious pathogens, machine learning methods and algorithms for data pre-treatment are used to translate Raman spectra of bacteria into a resistance prediction i.e. to predict pathogens and antibiotic resistances on the basis of the spectroscopically recorded data. For the compact microscope Medicars we use deep and machine learning techniques to translate multimodal image data into a tissue prediction for the detection of tumor margins. In smartphone microscopy, which is being researched by Rainer Heintzmann's team, image enhancement is achieved by means of deep learning procedures.

Where do the data sets come from that are currently mainly used? Can they be applied equally to all patients?

The data sets are generated within clinical studies, which we supervise from the beginning. The studies are still too small to exclude a gender bias, but we are working on the experimental design so that there is no gender bias in the training data set and we hope that the models will not generate any bias.

Does the automated analysis of medical control data also carry a risk? A loss of control?

Of course, every technology has risks, although these are manageable here. Artificial intelligence or machine learning processes only work well if the new test data is similar to the training data. We try to tackle this problem by creating the necessary similarity through standardization and model transfer in order to im- prove the predictions. There is a loss of control when the models are applied fully automatically. But in the medium term the models will only represent a second opinion, so there will be no loss of control.

Can physicians improve the learning systems? Is the procedure of AI applications comprehensible for them?

Physicians can increase the database or reduce the uncertainty of the metadata i.e. labels by pooling or voting, which leads to better models. The traceability of AI models is a major topic in current machine learning research Keyword "Explainable AI". The aim is to decipher these models in order to make it clearly understandable how mass-based learning methods and deep learning systems achieve their results.

Can AI be perfected to the point where it can eventually make better diagnoses than a human?

I think so, if the data is highly standardized. Another challenge is to demonstrate that improvement. This requires quite long clinical trials and is ethically problematic.

Could AI ever replace doctors instead of just to supporting them? For example, could operations be performed by AI-controlled robots at some point?

I don't think so, because there are many uncertainties in an operation that must be reacted to flexibly. This is not a prominent feature of current AI procedures. It's more likely that the surgical robots will do very specific things directly on the operator's instructions.

Will AI make medicine better?

In the long run, I think so. But first, it will make diagnostics more comparable and it will also allow data to be used not only sequentially, but in combination.

Artificial Intelligence, Machine Learning, Deep Learning

Decision making, problem solving, learning these are actions that we commonly associate with human thinking. We call their automation artificial intelligence (AI). An important part of AI is machine learning (ML). Scientists are researching algorithms andstatistical or mathematical methods with which computer systems can solve specific tasks.

For this purpose, machine learning methods construct a statistical-mathematical model from an example data set, the training data. On this basis, ML methods can make predictions or make decisions without having been explicitly programmed for it. ML techniques are used, for example, for spam detection in e-mail accounts, in image processing, and for the analysis of spectroscopic data. Deep learning is a method of machine learning that is similar to the way the human brain processes visual and other stimuli. Artificial neurons receive input, process it and pass it on to other neurons.Starting from a first, visible layer, the characteristics in the subsequent, hidden intermediate layers become increasingly abstract. The result is output in the last, again visible layer.

Making Tumor Tissue Visible with AI

Did the surgeon remove the entire tumor during surgery? In order to find out, researchers are combining optical methods with artificial intelligence (AI) and data pre-processing methods. AI is behind the compact Medicars microscope, for example, which enables rapid cancer diagnosis during surgery. Here, patterns and molecular details of a tissue sample irradiated with laser light are automatically evaluated and translated into classical images of standard diagnostics. Thus, tumor margins become visible.

"For this purpose, we train AI algorithms together with pathologists," explains Thomas Bocklitz. We take multimodal images of a tissue sample with our laser-based multi- modal microscope. In pathology, the tissue section is then embedded, stained, and an image of the HE- stained tissue section is taken (HE = haematoxylin-eosin). This enables the pathologist to recognize tumor tissue. Then we put the multimodal and the HE image side by side."

Based on the pathologist's analysis of the tissue structure and morphology, the research team teaches the algorithm which tissue is healthy and which is sick. "In this supervised approach, the algorithm learns to distinguish successive, healthy and diseased areas." With success: The accuracy of the predictions is more than 90 percent according to tests on a small group of patients.

Source: Leibniz IPHT

Continue reading here:
"Artificial Intelligence Will Make Medicine Better in the Long Run" - Biophotonics.World