guang000 / miccai_challenge_monuseg

This repo is for Miccai monuseg challenge, by Jimmy from aetherAI

7th place on Miccai - Multi Organ Nuclei Segementation

Chanllenge results: https://monuseg.grand-challenge.org/Results/

Slides: https://docs.google.com/presentation/d/1jS9YEs_KVBamoYdEZ0oSGUbIBQmr2htOz12dQLdf4Sk/edit?usp=sharing

Manuscript: https://drive.google.com/open?id=1S1apR4SV_aCiFbfLCaAkhh3EpJCfDCDu

---

Please install package below

pip install numba numexpr pygsheets oauth2client

First, setup your model hyper-parameter config in the monuconfig.py. We support backone: resnet50/101, densenet121/169 and inception-resnetv2, please set the model in BACKBONE.

class Config(object):
  NAME = "name your model"
  RPN_ANCHOR_SCALE = (2, 4, 6, 8, 10)
  BACKBONE = "resnet101"
  ...

Now support Path Aggregation Network and used as default. If you want to use original Mask-RCNN, please revise code in train.py when creating model

  # Create model
  model = modellib.MaskRCNN(mode="training", config=config,
                                model_dir=args.logs, is_PANet=False)

Then train the Mask-RCNN by

python train.py --weight imagenet --dataset dataset/ --logs logs/ --subset train

---

Already implemented features

• Path Aggregation Network
• Speed up data generator by
  • feed all data into memory first
  • apply Numba on utils.compute_overlap
  • rewrite utils.extract_boxex
  • revise some indexing code
• Support more pre-train model structure like DenseNet, Inception-Resnetv2
• Config and AJI results will be automatically recored on gsheets
• Speed up AJI code (implemented by 旻昇, 友誠)

TODO

• Synchronize Batch Normalization
• soft-NMS
• relation network
• Attetion on FPN