Review 2

[massich]

Reviewer 2 of ICPR 2016 submission 710 (Review4911)

Comments to the author

• [x] The size of the data set used in the paper seems to be too small compared with the other data sets mentioned in related work. This weakens the conclusions of the paper.
• [ ] Why leave-two-patient-out and not leave-one-patient out?
• [ ] Apparently the experimentation has been done as part of another work (Ref. 23), but the reference is not readable, so I could not judge this during my evaluation.
• [x] The paper states that it provides a public benchmark and then points to Ref. 22, which is again unreadable. In any case, that benchmark has already been provided in Ref. 22, so the wording by the authors is misleading.
• [ ] Review for grammatical errors and typos.

[massich]

2nd Comment

regarding Why leave-two-patient-out and not leave-one-patient out?, the paper refers to [12], but it might needs to be rephrased. Either to be explained, or to force the reader to go to [12].

All the experiments are evaluated in terms of \gls{se} and \gls{sp} (see Fig.\,\ref{fig:evaluation}) using the \gls{ltpocv} strategy, in line with \cite{Lemaintre2015miccaiOCT}.
Therefore, at each cross-validation iteration, a \gls{dme} and normal volumes are kept for testing, while the remaining volumes are used as training.
The \gls{se} evaluates the performance of the classifier with respect to the positive class, while the \gls{sp} evaluates its performance with respect to negative class.
Subsequently, no \gls{se} or \gls{sp} variance can be reported.
%However, \gls{ltpocv} strategy has been adopted despite this limitation due to the reduced size of the dataset.

[massich]

1st Comment

It was not clear that the second dataset from duke was only 45 volumes with 3 volume types, which is comparable to our 36 vol. 2 types. (I think I meade it more clear, and I clearly stated that we made the data public, which is +1 to us)