Artificial intelligence (AI) was a key topic at both MEDICA and the RSNA conference this year. But what are its applications in healthcare in general and radiology in particular? And what are the barriers? Dr. Michael Forsting, director of the Institute of Diagnostic and Interventional Radiology and Neuroradiology at Essen University Hospital in Germany talked to MedicalExpo e-magazine about his experiences with AI.
MedicalExpo e-magazine: What are the major challenges facing AI in healthcare?
Dr. Forsting: Artificial intelligence is not yet as popular in medicine as in other fields. But there is a good reason for that. There are not enough valid data to ensure optimal machine learning. AI algorithms are publicly available, but the systems can only be trained using absolutely valid data.
This problem applies to all tech companies—Google, IBM, GE, etc. All lack valid data. In addition, data sets are not always universal. Mammography screening is one example. Breast tissue in European populations is truly different from that of Asians. Consequently, AI systems trained with European data sets will not work in China or Japan.
ME e-mag: Medical image analysis is among the most promising applications for AI systems. Why is that?
Dr. Forsting: Image analysis, including radiology, is already digital. Furthermore, I believe AI affords radiology at least two major benefits. First, it reduces so-called satisfaction of search errors, typical in the discipline. For example, if a patient exhibits multiple sclerosis, the radiologist will focus on counting all such lesions, a long, tedious job. This thorough identification may cause the radiologist to miss a tumor.
“One of AI’s contributions to radiology is to extract more information from raw image data than ever before.”
The same applies to mammography or lung screening. AI can be trained to identify microcalcifications in the breast or small nodules in the lung, reducing such errors and improving radiology by at least 15 to 20 percent.
However, AI systems are now trained only with very specific diseases. In the lung screening example, AI might identify cancer cells while missing the tuberculosis detected by a radiologist. I believe this expertise is the radiologist’s real added value.
AI’s second contribution to radiology is to extract more information from raw image data than ever before. These systems analyze a much wider range of parameters than the human eye, multiplying prognostic information and enabling the identification of a large number of malignant diseases. MRI is an example. If a female patient has cancer cells in the uterus, AI will focus exclusively on these cells, analyzing more than 1000 parameters. It will predict with 97 percent accuracy whether the cancer has metastasized. This information not only suggests survival rates, but also patient response to different kinds of first, second and third line therapies.
ME e-mag: Will AI diagnoses break down disciplinary boundaries?
Dr. Forsting: My prediction is that major diagnostic disciplines, in particular radiology, pathology and laboratory medicine, will collaborate more closely in the future. We’ve learned from our analyses of uterine carcinoma and certain kinds of liver tumors that the system becomes more efficient if trained with both pathology and laboratory data. The fourth diagnostic tool, patient history, could be an AI-based questionnaire in which about 15 questions based on thousands of cases could determine which specialist a patient should see.
ME e-mag: Can AI lead to mistakes?
Dr. Forsting: AI systems are not totally error-free. Images do not always have a one-to-one match to a specific disease. Mimicry is possible. This means the radiologist is always responsible. But medical-legal questions will not stop the use of AI in diagnostic disciplines because these questions are solved thanks to very strict quality control programs.
“AI systems are not totally error-free. This means the radiologist is always responsible.”
ME e-mag: Does that mean that radiologists should not feel threatened by AI?
Dr. Forsting: Correct. At this stage, AI lightens radiologists’ jobs by freeing them from boring and time-consuming tasks. It offers them faster, more reliable diagnoses by reducing certain errors and increasing diagnostic possibilities. And there are no ethical issues involved because AI applications don’t incorporate specific patient data, even though they are trained using such data.
Obviously, this means that the practice of radiology will evolve. Radiologists will have to be trained accordingly, and will have to integrate this technology into their work. The same is true for pathology and laboratory medicine. We’ll also have to be creative when training AI systems to improve these disciplines and make diagnostic workups more efficient. Radiology will have to evolve and adapt. But the discipline is used to doing so—radiology is always on the move!