Fig. 4: Scale optimization and model performance on prostatectomy and biopsy specimens.
From: A comprehensive AI model development framework for consistent Gleason grading

a Model performances at different resolutions and the ground truth annotations. b Sensitivity, specificity, and F1 scores. The high-resolution model was selected for processing prostatectomy specimens because its F1 score is the highest for all labels. c Similarly, models at different resolutions were applied to biopsy images. d Considering the shape and size of the biopsy, a model at an extra-high resolution is more desirable. In particular, part of the biopsy was barely processed at lower resolution but was correctly identified as benign tissue at an extra-high resolution.