AI in radiology: reliable partner for diagnosing CT images

More patients, more examinations, more CT images – in radiology there is too much work for too few physicians. CT scans are evaluated in the shortest possible time, which leads to anomalies being overlooked. Artificial intelligence, on the other hand, works with constant speed and performance, which is why radiological routine increasingly relies on its support.

In this MEDICA-tradefair.com interview, Ivo Driesser explains what is possible with the intelligent software assistant AI-Rad Companion Chest CT from Siemens Healthineers, why Artificial Intelligence does not make a diagnosis and what the digital patient is all about.

Mr. Driesser, what does a radiological examination normally look like?

Ivo Driesser: Normally, the patient comes to the hospital for a certain reason, for example to examine the lungs. He is registered and taken to the CT. After the thorax has been scanned, the images are either sent directly to a Picture Archiving and Communication System (PACS), where the physician can analyze them. Another option is to first send the images to a workstation where MTAs or physicians convert the scans into 3D images or take measurements before they are sent to PACS for diagnosis. The data set of a single CT scan can consist of fifty to a thousand images that the physician manually goes through to look for abnormalities.

If we look at today's developments, we see that there are more and more patients and therefore more and more work. However, the number of physicians is not growing proportionally. As a result, physicians must work longer or scan faster. Studies show: If the physician has fifty percent less time to evaluate an image, the error rate increases by seventeen percent. Abnormalities are overlooked because the physician is tired or does not have enough time. And this is exactly where we have to start.

With the AI-Rad Companion Chest CT you can solve these problems. In what way?

Driesser: We want to help the physician become faster and more accurate by replacing manual image analysis with Artificial Intelligence analysis. The AI-Rad Companion Chest CT is an intelligent software assistant for CT images in the chest area - heart, aorta, lung and spine. Further applications for the AI-Rad Companion platform are planned. The algorithms can use the scans to identify the organs, perform measurements and combine several images to form a 3D image. The automatic measurement saves the physician five to ten minutes per patient.

If a patient comes to the hospital with the requirement that the lung be examined, this is exactly what is done. For example, the physician checks whether there are round foci in the lung or not. Artificial Intelligence, on the other hand, automatically examines not only the lung but the entire thorax and could, for example, detect calcifications of the coronary vessels without specifically asking for them. This is Incidental Findings, which the AI-Rad Companion provides the physician with and which can be very helpful. If, for example, the patient had lung cancer, the problem with the coronary vessels would first have to be solved before cancer treatment could begin.

The AI-Rad Companion compares the patient's values with reference values stored in the software. If it finds deviations or abnormalities, it marks them in different colors. In addition, it compiles a pictogram of the thorax with all organs based on the measured values. For example, if the heart is marked red, abnormalities were found there. The algorithms do not make a diagnosis, but show the physician which site he should look at more closely.

What requirements must be fulfilled for the implementation and use of the application?

Driesser: The AI-Rad Companion runs on our teamplay cloud platform. The advantage of this solution is that the hospital does not necessarily have to invest in additional or new computing power.  As soon as the IT requirements in the hospital are met, we install software there that securely sends the images from the CT scanner to the AI-Rad Companion, processes them in the cloud and forwards them to the physician in the PACS. Once the customer has an account for the platform, they can use new applications and services from Siemens Healthineers without having to invest in new hardware or come to the customer to set them up. It's also important that the images taken have a certain quality. Certain criteria must be met, such as the thickness of the layer, so that the algorithm can evaluate the images. Before the images are processed, the system performs a data check and checks whether the CT scans have the necessary quality.

You have just mentioned that such technologies are increasing more and more. Can the AI dispute the job of the human physician? And what tasks still have to be performed by radiologists?

Driesser: I have spoken to many radiologists who are not afraid for their job. The physician finds the support he gets from the AI very helpful. He simply sees that the algorithms always do the same thing without getting tired. If, on the other hand, the physician diagnoses images all day, he gets tired and the probability increases that he overlooks conspicuous features. Of course, he is happy to help with this task. Young radiologists with little experience and an untrained eye are helped by the data provided by the algorithms and additional measurements. For example, the AI can also be used to train skills or expand one's own knowledge.

As already mentioned, the scans must have a certain quality. Images that do not withstand the system's data check or that the algorithm does not know are not processed. The physician has the possibility to see the results before they are saved. So the human physician always has the last word. Our software assistant is just a companion – a partner – of the physician.

What other possibilities of AI could you imagine for the medicine of the future?

Driesser: Data is collected from a patient inside and outside the hospital - in pathology, radiology or cardiology, but also from our own Smartwatch, for example. The goal is to use all this data in a meaningful way. We are already working on this. Before a patient starts a therapy, all data should be collected and processed by AI to create a digital twin of the patient. If the physician decides on a specific therapy, it would first be tested virtually. This would allow results to be estimated in advance and risks to be minimized. The more I know about a patient, the more I can predict.