Fig. 1: Follow-up period and reference standard—comparing radiologists double reading to AI systems. | npj Digital Medicine

Fig. 1: Follow-up period and reference standard—comparing radiologists double reading to AI systems.

From: Simulating mismatch between calibration and target population in AI for mammography the retrospective VAIB study

Fig. 1

Impact of different follow-up times as the basis for the reference standard for AI performance compared to radiologists double reading for the three AI systems. As the follow-up period is shortened the bias in favor of the radiologist increases (the ‘x’ denoting radiologist performance is further away from the ROC curve of AI). Even if the overall accuracy is lower when using a 3-year (purple) follow-up period since cancer diagnosis is more distant from the time of the mammography, it is the shortest timeframe to avoid bias in favor of radiologist compared to AI detection.

Back to article page