Supplementary Figure 3: Determination of the accuracy of CDeep3M mitochondria segmentation, based on one FIB–SEM and one SBEM dataset.

As noted in Lucchi et al., human ‘ground truth’ segmentations are typically inaccurate around the borders of an object. a, An exclusion zone of 1–2 voxels can compensate for this effect and avoid erroneously assigning those pixels. b, We determined the accuracy of CDeep3M prediction on an FIB–SEM hippocampal dataset, using the same metrics as described in Ref. ⁸. CDeep3M outperformed the three-class CRF in all metrics (Jaccard: CDeep3M: 0.8361 versus 3C-CFR: 0.741; two-voxel exclusion zone: CDeep3M: 0.9266, 3C-CFR: 0.85; five-voxel exclusion zone: CDeep3M: 0.9437, 3C-CFR: ~0.92). Both the Jaccard index and the F1 value (the harmonic mean of precision and recall) increase once the erroneously missing object boundaries in the human segmentation are masked by the exclusion zone. The remaining error was largely caused by a single large object in the test data, which resembled the appearance of a mitochondrion and which was absent from the training data. c, d, Similarly, we used the SBEM data shown in c (scale bars: left, 500 nm; right, 200 nm) to compare computer versus repeated human performance. d, The consensus of three ‘ground truth’ segmentations of expert human annotators was used to determine the performance of CDeep3M and compare the individual performance of each human annotator to the consensus. CDeep3M performed similar to the human experts (exclusion zone of 1 voxel; Jaccard index: CDeep3M: 0.954, humans (mean): 0.983; F1 value: CDeep3M: 0.976, humans (mean): 0.966).