I congratulate the authors on a great piece of research. I believe that the investigations conducted in this work are critical to unlocking the full potential of deep learning tools for image segmentation and improving accessibility. A little more clarity could be provided in some methodology choices.

L127 It is unclear why these were scaled to 256x256x256, does it not increase computational time with no benefit as upscaling should not produce more information? Please clarify this step.

L154 Why were the median accuracy networks used?

L638 The Table 2 description is hard to read. Why is it not included in the table column names that F statistics are presented?

L623 Figure 7 vertical axes all say PC1, should it not be compared to PC2 and PC3 AI shapes for C, D and E, F. Respectively.

L623 Figure 7 B D F are not marked with triangles for the original data A C E are, the data type legend does not include the point style. Please make these consistent.

The percentage of variation covered by each PC could be included somewhere. Either text or the figure.

It is unclear why this type of augmentation was chosen. Why rotated volume and not filtered volume instead, why not both or why not something else? There is no reason provided as to why it was rotated 5 times. Presumably, it would work better if more rotations were included. But this was not presented or discussed. No sensitivity tests were carried out on the augmentation to justify the number of rotations or types of augmentation. Some justification here would be appropriate.