IRA - Intelligent Radiological Assistant

IRA - Intelligent Radiological Assistant

In the evaluation of imaging procedures (X-ray, CT, MRI), many assistance systems already exist to visually prepare the images for diagnosis. Software tools support physicians by segmenting and annotating regions of interest in medical images and diagnosing diseases. In recent years in particular, the analysis of medical image data using artificial intelligence has made great strides.

In order to increase acceptance among physicians, diagnoses should be made comprehensible. However, the greatest advantage of frequently used neural networks is at the same time a great disadvantage: Due to their complexity, they allow classifications to be made for difficult questions, but the underlying decision-making process can hardly be traced. However, the comprehensibility can be increased by novel methods through so-called "Explainable" or "Reasonable" AI. Here, especially for image data, a modified version of the image is usually output, in which specific areas are marked that are of high importance for the network for a certain diagnosis/classification.

At SBI, existing algorithms of "Explainable" AI are applied to medical diagnostic tools, and new methods are developed specifically for this use case.

In modern medicine, there is hardly a process in which a computer is not used in one way or another. The range of application of computers and special software ranges from simple auxiliary programs to data management for entire clinics. A multitude of assistance systems help doctors and hospital staff to link enormous amounts of data through increasingly powerful IT systems.
A few years ago, the development of such assistance systems was mainly driven by universities and other research institutions. While a few years ago the development of such assistance systems was mainly driven by universities and other research institutions, this development has now also reached the private sector. Cooperative ventures between companies and research institutions are also appearing more frequently. There is a market for intelligent programs that support physicians in their work, which will continue to grow in the coming years thanks to ever better technology and advancing digitization.
In the evaluation of imaging procedures (X-ray, CT, MRI), many assistance systems already exist to visually prepare the images for human diagnosis. Software tools support physicians by segmenting and annotating regions of interest in medical image data as well as diagnosing diseases. In recent years, the analysis of medical image data with the aid of artificial intelligence (especially Deep Learning: Convolutional Neural Networks, CNNs) has made great progress. On the one hand, there are increasingly better CNNs that solve the classification of image data. On the other hand, datasets are increasingly being published in medicine, which can be used to train neural networks. By using these data sets, it is already possible to train CNNs that can achieve a comparable quality of human experts in their diagnoses.
A diagnosis made exclusively by a CNN cannot be equated with a human diagnosis from both an ethical and a legal point of view, even if the probability of error of the diagnoses made is statistically smaller. This is because when a CNN makes a finding for a disease, it is unclear on what basis it made that decision. Because of this uncertainty, CNNs will not replace physicians in the near future.
However, the technology can be used as an assistance system to increase the quality and efficiency of human diagnosis. Physicians spend a lot of time evaluating radiological images. A system that can sort out negative findings with a high degree of accuracy will significantly reduce the workload of medical staff.

 

Description of the "Reasonable IRA" subproject

With the help of the final assistance system, the attending physician should be able to confirm or reject findings as accurately and time-efficiently as possible. This can be enabled, for example, by a visual output that highlights critical or important areas in the radiological image. The main task of the SBI is the implementation of these methods of Reasonable AI into the IRA workflow. For this purpose, existing algorithms on the classifiers of PLANET AI and the corresponding data. Subsequently, the most suitable explainability algorithms will be specialized on the classifiers developed by PLANET AI. This can be done by combining of existing algorithms as well as the development of new methods.