Computer-aided tumor cell fraction (TCF) estimation by medical students, residents, and pathologists improves inter-observer agreement while highlighting the risk of automation bias.

The potential for computer-aided diagnostics (CAD) to augment learning outcomes for medical education in histopathology are vast. However, over-reliance in CAD tools could lead to important errors if the output is inaccurate, such as automation bias. Here, we analyze the influence of a CAD tool for tumor cell fraction (TCF) estimation, named TCFCAD, on medical students with limited histopathological expertise, compared to pathology residents and practicing pathologists. Participants were third-year medical students from the University of Bern (n = 18), residents and pathologists from the European Masters in Molecular Pathology (EMMP; n = 23), and 63 practicing pathologists stratified by < 20 or > 20 years of professional experience (n = 28 and 32, respectively). Each group evaluated 10 colorectal cancer regions of interest (ROI) with and without TCFCAD assistance. The ground truth (GT) was evaluated by a laborious count of all tumor and non-tumor cells per ROI. All groups demonstrated reduced variability in TCF scores with TCFCAD assistance. The standard deviation (SD) of the TCF scores compared to the GT values before and after TCFCAD assistance were 9.09 to 4.95 for medical students, 9.93 to 5.55 for EMMP, and 9.98 to 5.69 and 9.9 to 6.13 for pathologists with < 20 and > 20 years of experience. For one image, both medical students and EMMP participants' scores swayed to the TCFCAD's output despite (1) tool output being inaccurate and (2) participants' original scores without TCFCAD assistance being closer to the GT value. This confirmed a trend of medical students following the tool's recommendation across all ROIs. All groups benefitted most from TCFCAD output for a tumor with marked inflammatory infiltrates. Although TCFCAD improved interobserver agreement, independently of professional experience, there is a demonstrated danger in over-relying on its output. This important bias must be addressed when training students and pathologists on the use of CAD tools.