Figure 1.
Synapses, detected in the test dataset.
A: A 3D view of the synapses, detected in the test dataset. The detected synapses are shown in green. The central slice of the raw data is also shown for illustration (the shape distortion is caused by elastic stack registration). The data volume was downsampled by a factor of 10 in the x and y dimensions to show the full volume. B, C, D, E: More detailed synapse examples as an image series. Scale bars: 450 nm, every second slice is shown (distance between consecutive images is 90 nm).
Figure 2.
A - distribution of false negative errors as a function of the synapse average cross section area and continuity in z. For each bin of the 2D histogram its count is proportional to the radius of the displayed circle. Cross section area was measured in pixels and averaged across 5 central slices. B, C - serial sections of false negative detections. D – erroneous algorithm segmentation. All scale bars – 450 nm.
Figure 3.
False negative errors against synapse size and perforation.
Left: distribution of false negative errors as a function of synapse size (see text for details on size estimation). Right: distribution of false negative detections as a function of the number of slices, where the synapse is perforated.
Figure 4.
Examples of false positive detections.
All scale bars – 450 nm, every second slice is shown (distance between consecutive images is 90 nm).
Figure 5.
The proposed synapse detection pipeline.
Left to right: raw data with 3 synapses, shown in green circles; probability map of the synapse class; detection results; graph cut segmentation results; object classification results, with positively classified objects shown in green and the negatively classified object in red.
Figure 6.
Benefits of 3D processing with upsampling.
A: raw data, with three synapses in green circles. B: probability map of the synapse class after classification with 2D features – low precision of the prediction. C: same for classification with 3D features without upsampling – low recall. D: same for classification with 3D features and upsampling by 2 along the z axis. The remaining false positives are filtered out by the object classification step.
Table 1.
Pixel features.
Table 2.
Object features.