To map the learning progress of less experienced readers in prostate MRI segmentation.

Fifty bi-parametric prostate MRI scans were randomly selected from the Göteborg Prostate Cancer Screening 2 Trial. The MRI scans included T₂-weighted imaging according to the PIRADS v2.1 guidelines. Nine annotators with a varying degree of segmentation experience were involved: 1 expert radiologist (“expert”, >1000 prostate-MRI segmentations), 2 experienced radiology residents (“experienced residents”, 100-200 segmentations), 2 inexperienced radiology residents (“inexperienced residents”, no segmentation experience) and 4 unexperienced medical students (“novices”, no segmentation experience). The task was to segment the whole prostate gland excluding the seminal vesicles. The expert segmentations were used as ground truth. For all other readers 50 MRI scans were divided into 3 rounds. After each round a one-on-one...

...read more

At baseline (first round) the DCS of residents regardless of prior experience was 0.91±0.19. For the subsequent rounds there was no significant change over time and DCS of the last round was 0.91±0.03. Novices produced a significantly lower DCS at baseline of 0.76±0.09 and achieved a higher DCS of 0.84±0.03 in the last round, i.e., in the last round the DCS of residents was higher. All other metrics showed a similar trend.

The residents that participated in this study provided segmentation data close to the expert’s even at baseline. The novices showed inferior agreement but systematic feedback improved the agreement but not to the same level as the residents. This study highlights that care should be taken when using novice annotators and they should be offered more extensive training and education than residents prior to segmentation work.