The Expert.

Top Line: AI is looking to win back prostate cancer pathology.

The Study: Slides from 8571 biopsies from 1289 men in the Swedish STHLM-3 prostate screening study were randomly selected and digitized. A single pathologist assigned grade groups, delineated malignant areas, and measured the length of core involvement. Two additional test datasets were also assembled from outside the STHLM-3 program. Here’s how the model worked. First, image segmentation detected tissue from background followed by annotation of malignant tissue. The full slide image was then broken down into smaller overlapping patches. These segmented and annotated patches were then used to train the deep neural network (DNN). The DNN had two ensembles that first classified patches as benign or malignant and then determined the primary Gleason pattern (3-5). An “ensemble” is a collection of multiple sub-models that analyze the same data. Their predictions are averaged to slightly improve the accuracy of the overall model. To train tumor grading, they used only biopsies with the same primary and secondary Gleason patterns (e.g., 3+3 or 4+4). The DNN was then trained to put these pieces together to assign a single grade group to the core and to measure the length of the core involved. The error rate for malignant versus benign was <1% in the test set and <2% in the external validation set. The kappa statistic for grade group agreement was within the range of a group of 23 expert pathologists reviewing the same biopsies, although it didn’t significantly outperform these experts.

TBL: AI is learning to use the trees to see the forest in prostate cancer grade grouping. | Strom, Lancet Oncol 2020