Paper updates

paper/paper.tex

@@ -291,10 +291,14 @@ \section{Conclusion and discussion}
 
 One notable question is if we might get different results if users were aware of the larger ``pen size''. Might they make fewer errors and redundancies? This would also avoid the need for our dilation algorithm to utilize a non-traditional approach in trying to avoid introducing new errors. There's the worry that this approach could create unrealistic results.
 
+There was some cases where moderate amounts of dilation actually introduced worse results (see figure \ref{fig:yuanxia_median_dsc}). There was still an increase for radius 4 pixel dilation, however. It's unclear why the algorithm acts this way. The segmentation ends up being fairly different, but there's no obvious explanations for this and it's something that would need further study. This only occurred in the results from Li's study, so perhaps the time pressures are related to this. An analysis could be done on if these odd dilation results correlate with user errors or the like, since the time pressure introduced more errors.
+
 The OneCut algorithm that we tested proved interesting. In general, it should probably be regarded as little more than a side consideration in this study. There's the big issue that users were not trained with this algorithm and thus did not realize that it is much more sensitive to how much information is provided. As well, it just plain works very poorly on undilated label images. It was clearly not created with so few seeds in mind and the Rau and Li studies were not designed with that requirement. Hence, the provided user data is insufficient to well test it, especially when we have so many cases of no segmentation occuring, which essentially ends up statistically drowning out relevant results. However, we can at least conclude that the OneCut algorithm is extremely sensitive to the amount of annotations and thus recommend that its users provide many, thick strokes if they wish to get meaningful results.
 
 The fact that semi-automatic image segmentation algorithms seem sensitive the quantity of seed points (which is really what dilating the strokes is doing) also serves to partially explain the results of strokes versus points. Stroke annotations naturally provide more seed points than point annotations, as dilation does. That said, strokes are much more efficient than points because they provide more \textit{useful} annotations (due to greater range of area). And as we've shown, thicker strokes perform better still.
 
+This does also mean that in comparing the results of studies that utilize a seed-based approach for semi-automatic image segmentation, we must consider the width of the strokes used. As well, the width should also be consistent within the study in order to have truly comparable results.
+
 The code and data for this study are available online at \url{https://github.com/KatrinaHoffert/stroke-radius-segmentation} \cite{repo}.
 
 \bibliographystyle{./IEEEtran/bibtex/IEEEtran}