Tourassi warns that the use of AI tools is not meant to replace radiologists but produce second points-of-reference that are free of contextual bias, thereby reducing the occurrence of errors associated with it. She adds that though promising, the results of the study require testing on larger sets of data.

“Our initial findings were based on a limited imaging dataset. The next step is to expand our experiment to other clinical sites with a larger imaging dataset and more radiologist participation. This larger experiment will confirm the robustness of our initial study findings. Further, we will proceed with developing a dynamically adaptive decision support system which integrates imaging and radiologists' gaze data to reduce the risk of human error in medical image diagnosis.”

Imagery reflected a range of cases typically found in clinical settings, consisting of positive and negative diagnoses as well as cases that mimicked the signs of cancer but were declared benign. No prior knowledge of the findings for each was disclosed to participants. Deviation in the context of different image categories was also calculated.

Though tested for mammographic interpretation, the same experimental design and algorithm can be applied to diagnostic interpretations involving different imaging modalities and diseases.

Research was supported by ORNL’s Laboratory Directed Research and Development Program with the work relying on the resources of the Oak Ridge Leadership Computing Facility, a DOE Office of Science user facility.

The findings were published in the Journal of Medical Imaging.

Back to HCB News

Women's Health Homepage