What is it about?
We try to achieve state-of-art results with the BreakHist dataset! Considering non-reported specificity or ROC-AUC metrics in literature, we aim to achieve a good (less) false positive rate! We use pre-trained and finetuned mobile networks, such as the Resnet family or MobileNet itself.
We published (inaccessibly) our first draft of the work! Congratulations to my dear team members!
Soon, you can reach our report on my website, here!.
Future Discussions:
- We really want to utilize superpixels on this dataset somehow, it can be a good on-air-augmentation methodology, as we saw patching is almost catching up with a non-augmentation methodology.
- We want to apply better feature selection on the ML branch of work.
- We desire to provide some advantageous RetinaNet-like architecture, given some arbitrary bounding boxes.