This contains our solutions for 2018 data science bowl hold by Kaggle. The task is to do nuclei segmentation across multiple kinds of cell images.
Our team ranked 407 out of ~3800 teams, with final score 0.436. We first tried basic image segmentation without any deep learning. Then tried "semantic segmentation" SegNet, a Fully-Convolutional Neural Network. Finally we implemented Mask-RCNN, and got the best score among our tryings. We did not try Unet, which may yield a higher score.
- Basic image processing method using watershed (Visualize_data_and_basic_segmentation.ipynb)
- Implementation of Mask-RCNN (Mask_RCNN.ipynb)
- Source code of segmentation functions (segmentationFunctions.py)
- Python 3
- Common python libraries including numpy, cv2, skimage, etc.
- Download Mask-RCNN from https://github.com/matterport/Mask_RCNN, and add the folder to system path for Mask_RCNN
- Download nuclei images from https://www.kaggle.com/c/data-science-bowl-2018/, and change corresponding paths in the scripts