Figure 7: 3D semantic segmentation.

(a) To adapt Merlin for segmentation, we add a decoder and skip connections between the Merlin encoder and decoder. (b) We compare model variations using average Dice score across 20 organs that appear in abdominal CT. We compare performance of models trained using 100% of training cases and also simulate the data scarce regime with 10% of training cases. (c) We report Dice scores for 20 organs across 4 model variations using 100% of training cases. (d) We qualitatively compare segmentations between the ground truth labels, a model with the Merlin architecture that has been randomly initialized, and Merlin. The red arrows indicate mistakes made by the model relative to the ground truth. Each row is a different patient sampled from the Total Segmentator⁵¹ test set.