| dc.description.abstract | In this master thesis we have adapted and implemented Mask R-CNN to the task of detecting and localizing nuclei in medical imaging.
Mask R-CNN, which does instance segmentation, was chosen as the architecture to implement, based on a literature review.
Our best Mask R-CNN model achieved a F1-score of 0.385 on the validation set and 0.460 on the test set.
This thesis was inspired by and a part of the Kaggle 2018 Data Science Bowl.
We did not complete our implementation in time to enter the competition, but if we had, this score would have lead to 291st place out of 738 participating individuals or teams in the Kaggle competition.
The differences between our implementation and the second-placed implementation, which also implemented Mask R-CNN, were mainly a different backbone and heavy data augmentation.
This shows that our approach was competitive, and with modifications could have been competing for the top placements in the competition.
The code is available at: https://github.com/jolohan/detectron2.git | |
