To classify exudates (spots) representing diabetic retinopathy (DR) in fundus images, I've taken the following two approaches:
1) The sum is taken of all objects for each features. These sums are used as image level features. This classifier would then produce a single label for an unknown image, either pathological or healthy.
2) Every object in each image has features extracted that are used for training. The label assigned to each object feature set is the DR grade for that image. So, if an image is pathological (1), then all features extracted from that image are assigned label 1, and conversely for healthy images. This classifier would then produce a set of labels for an unknown object, classifying each object contained in the unknown image. We could then say if any object in the image is pathological, then the image is pathological.
I'm receiving better results for method 2 than for method 1. Is this an acceptable measure of classification performance to be used in publishable research?
Use both first and second calculated features in a matrix and feed them into the training algorithm, you may have better result.
see attached.
And also if you add some more context features like the number of spot in the spot near neighborhood, you second algorithm would work fine
I think the second approach is more appropriate and I could imagine it would be more reliable. Let's assume that there are 100 objects in an image and 99 of them are negative (not pathological). Using method 1, the sums of the features will be dominated by the features of the negative objects while the one positive objective will probably have a comparatively small contribution. Therefore, method 1 will probably classify this image incorrectly, but method 2 will likely classify it correctly if it can find the 1 pathological object. Method 2 seems to be the more robust option.
However, I think the training stage in method 2 can be performed better. It would be ideal if each object in every image would have its own corresponding label (pathological or not). Then, each object would be a separate training sample instead of each image.
since the question is about performance measure, in a publishable research you'd need to define what do you mean by "first method gives better results than the second". You can provide reasoning similar to one provided here by Matthew G Keffalas, but also you'd want to compare the classification accuracy (sensitivity, specificity, etc.) on real or synthesized DR data. Real data are always preferable and validity of your comparison will be viewed in terms of statisitcal significance of better classification by one method compared to the other.
- Badges
- Science topic
- Similar topics
- Correction