Expert Review of Algorithmic Mitotic Count Predictions Ensures High Reliability

Computer-assisted mitotic count using a deep learning–based algorithm improves interobserver reproducibility and accuracy

“… 29 , 33 A recent pilot study has also reported a significant time improvement when using MC heat maps and MF candidate visualizations as a guide for a pathologist. 15 In the present study, participants noted an enormous time improvement between stages 1 and 2 due to omission of manual MC-ROI selection. However, we decided not to measure labor intensity of the 3 stages systematically, as our focus was to evaluate the potential benefits, limitations, and requirements of these software tools.…”

“…However, we highlight that software for routine application should allow the pathologist to modify the MC-ROI location if the preselected MC-ROI is considered inappropriate due to algorithmic failure. While it was beyond the scope of the present study to investigate approaches for software-pathologist interaction on MC-ROI selection, we propose that computer-assisted MC tools include one of the following features: (1) restriction of the algorithmic predictions to the actual tumor area, which can be delineated manually by a pathologist 7 or segmented automatically by means of deep learning–based algorithms; 22 (2) heat map visualization of mitotic density for quick manual selection and correction of the MC-ROI; 15 (3) proposal of the top 3 to 5 hotspots areas from which pathologists can choose. Another aspect that needs to be considered for future implementation of computer-assisted approaches is the shape of the MC-ROI.…”

“…A recent pilot study has shown that expert review, that is, computer assistance as opposed to unverified algorithmic predictions, is necessary to ensure high reliability of the MC in cases that have a presumed marked domain shift (and thus higher rate of algorithmic errors). 15 Therefore, we consider computer-assisted approaches with verification by a trained pathologist to be highly beneficial, at least until more experience and progress is gained through research and routine use of these applications. Access to intermediate results of the algorithmic pipeline (such as visualization of MF predictions) may allow a higher degree of comprehensibility by the reviewing pathologist and should therefore be encouraged.…”