Variable AI predictions predominantly occur at object boundaries during segmentation. (A–E) 3D test data acquired via FIB-SEM imaging, showing an (A) overview and (B) inset 2D slice from the dataset with (C) corresponding predictions generated by a random forest classifier trained to detect membranes, alongside (D) 2D binary segmentation and (E) 3D surfaces based on the membrane predictions. (F) 2D slices depicting raw membrane predictions for the test dataset (from Fig. 1 B) with values (V) shown as percentages of their total dynamic range (0–1), made using six different machine learning algorithms (also provided as 3D renders in Fig. S2 A); RF, J48, MLP (multilayer perceptron), DT (decision table), JRip, and PART (projective adaptive resonance theory). (G) Visual depiction of the absolute difference in voxel values (ΔV) between each prediction from F, showing the average value difference for each two-way value comparison as a percentage of the total dynamic range (0–1). Scale bars; A, 1,000 nm; B–G, 200 nm.