AI makes MRI diagnoses as accurate as doctors. But 4 times faster

Researchers from FastMRI have presented an AI model that can analyze and diagnose MRI scans. In most cases, it is no worse than living specialists.

Radiological experts have proven that artificial intelligence can evaluate MRI results, make diagnoses, and recommend treatment. At the same time, the model does this as well as ordinary doctors. In a blind comparison, experts were unable to distinguish between the findings of the AI ​​and those of the doctors. The system works four times faster than a live specialist, so it can reduce waiting times and costs for additional examinations.

The FastMRI team based their model on the assumption that some of the collected MRI data is redundant and unnecessary for conclusions. This means that a well-trained machine learning system can make its own conclusions about which data is important for further conclusions and which are not. After that, the scientists trained the model on a large amount of data, since the MRI scans are very orderly and predictable.

The study showed that there were no significant differences in the assessments of specialists and AI. They found the same abnormalities and pathologies regardless of who drew these conclusions. All researchers rated the findings from AI as better than traditional ones. Five out of six radiographers were unable to correctly identify which images were processed by the AI.

Engineers note that there can be differences between the conclusions of AI and specialists only if there is redundant data or “noise” in the original data. In this case, only a radiologist can draw the right conclusions, so for now the model’s conclusions are being checked by a living specialist.