Fig. 2

Training and inference protocol, and example training labels for RR segmentation. A Training and inference protocol used in this work. The training datasets were split into 5 disjoint sets, with a UNet model trained on each combination of 4 sets, testing on the final test set each time. After the training and testing on the primary dataset is complete, this gives us 5 UNet models that are used in ensemble for inference, with the output logits being averaged and thresholded to gain the final segmentation. B  A sample of the raw images (left) along with the expert annotations for Rods and Rings used for training. It is worth noting that although these labels are provided by a highly skilled expert in the field, they are by no means an “absolute” ground truth; and contain some implicit biases and variations between the images. Scale bars represent 30 μm