Deep Learning: How artificial neural networks can support diagnostics

The use of artificial intelligence and deep learning in medical diagnostics is growing rapidly. Ubotica Technologies won an open call as part of the FED4SAE innovation measure backed by fortiss GmbH: Ubotica's neural network is based on deep learning and detects the presence of diabetic retinopathy in retinal images. The eye disease is a diabetes complication that can ultimately lead to blindness.

Dr. Holger Pfeifer heads the Software Dependability Field of Competence at fortiss, the Research Institute of the Free State of Bavaria for Software-Intensive Systems. In this MEDICA-tradefair.com interview, he details the research contribution made by fortiss, talks about the project successes, and reveals the obstacles researchers must continue to overcome in adopting deep learning systems.

Dr. Pfeifer, how is diabetic retinopathy normally being detected and how does your innovative diagnostic method work?

Dr. Holger Pfeifer: Physicians currently use a retinal camera to photograph the interior surface of the eye, including the retina. They use these images and certain indicators that suggest damage to make a diagnosis. Ubotica Technologies has developed an artificial intelligence (AI)-based solution designed to automate this assessment. The idea is that we feed the camera images to a neural network that subsequently makes a classification: it detects if diabetic retinopathy indicators exist or not. It is a screening tool that helps doctors to make this initial classification.

At fortiss, we create software tools that assess the quality of the AI technology. Our goal is to assure the accuracy of these AI-based solutions, making sure they detect the disease indicators reliably and don’t give false-positive results when there are no indicators. Another aim is to study and understand why certain cases render a "yes" or "no" result – and what are the decisive characteristics in this setting?

In this specific application, Ubotica wanted to integrate the AI process directly into the camera. Usually, AI research is computationally expensive and often requires powerful computers or cloud applications. However, the notion of sending this data to a cloud to conduct an analysis is precarious as it pertains to data protection and privacy. That is why the idea was to use chip technology to implement the AI ​​model directly on the chip and integrate it in the camera to facilitate an analysis during the eye exam.

This is a critical use case, and we want to support Ubotica with our Neural Network Dependability Kit (NNDK), thus enabling them to better understand their model in terms of its accuracy and size. After all, the model must be small enough to fit the chip and still be able to make fast calculations. Our NNDK software tool can facilitate all these questions.

We managed to reduce the size of the original chip significantly and were able to determine that the reduced size did not have a significant impact on accuracy when classifying the retinal images.

Lastly, we benefitted from yet another NNDK feature: We can store how the AI technology responds to an image and filter out similar images of other patients from our data pool that most closely match the image of the current patient. The physician sees that the AI technology reaches a similar decision based on the resemblance to other cases. The images were previously classified by experts and categorized as disease examples. The eye doctor gets additional clinical decision support because the AI technology substantiates its classification.

How far along are you in your research and when will the early detection system be used in practical application?

Pfeifer: We developed a prototype of the application as part of the project to test the feasibility and viability of the solution. You must overcome major challenges before you can bring a new medical device to market - though we did not analyze the regulatory requirements within the scope of this project. Having said that, fortiss definitely considers these issues in other applications that pertain to autonomous driving, for example.

What are the key challenges of machine learning?

Pfeifer: Artificial neural networks are always riddled with uncertainty since they can only learn from provided data. Yet how do they respond to new, unknown samples? Training deep learning neural networks is very challenging and complex: you need a lot of training data, which must often be pre-classified by experts. The processes are also not fully transparent since AI-technology performs complex mathematical calculations which are difficult to reconstruct, begging the question: What does AI do and how does it really work? Why does it end up with these results? If you understand which parts of the model are responsible for certain aspects in an image and realize how these parts respond to certain indicators, then you can comprehend the robustness of the neural network when the images are slightly corrupted, for example.

We are concerned with the traceability and trustworthiness of such applications. Fortiss studies key aspects such as robustness, accuracy, and interpretability as essential properties of trustworthy classification. We collaborate with companies that develop interesting AI applications. Our research helps them improve the quality of these applications as it pertains to traceability and trustworthiness.

What role will deep learning play in medicine in the future?

Pfeifer: Deep learning will certainly play a fundamental role in this area. In medicine, we must often detect tiny details, especially in the case of imaging technology to diagnose and treat diseases. AI technology is a great tool for medical image analysis.

AI can also do great things with already available data. Just think of all the data and medical findings and insights, which are often only available by word-of-mouth or in paper format: This data is a hidden treasure that has yet to be found. It doesn’t necessarily have to involve neural networks since you might also be able to use other AI technologies. AI ​​can help prepare this data for use, making it possible to benefit from this knowledge in diagnostic decision making and treatments.