report of kaggle competition - indico.mpi-cbg.de · report of kaggle competition (gahobe-gayathri,...

9
Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon ([email protected]) 5th Nov. 2018 1 Scientific Computing Facility@MPI-CBG

Upload: others

Post on 22-May-2020

16 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Report of Kaggle Competition(GaHoBe-Gayathri, HongKee, Benoit)

Scientific Computing FacilityHongKee Moon ([email protected])5th Nov. 2018

1 Scientific Computing Facility@MPI-CBG

Page 2: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

2018 Data Science Bowl

• Mission: Nuclei detection automation1

• By using Deep-learning API• Python preferable

• Our approach• Keras (tensorFlow backend) / U-Net• Fiji• Parameter optimization

1 https://www.kaggle.com/c/data-science-bowl-2018

2 Scientific Computing Facility@MPI-CBG

Page 3: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Pipeline

1. Training a model - U-net2. Testing the model on the problem set3. Submission and post the score

Requirements1. python: numpy, skimage, keras2. fiji: image processing3. keras: deep learning

3 Scientific Computing Facility@MPI-CBG

Page 4: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Training a model

1. input: 512x512x3 8bit, output: 512x512 boolean2. preprocessing resize, boundary weights, masks3. loss function, accuracy function for optimizer4. U-net networks5. fitting the model

6. check IoU2 results2 https://www.pyimagesearch.com/2016/11/07/intersection-over-union-iou-for-object-detection/

4 Scientific Computing Facility@MPI-CBG

Page 5: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Testing the model on the problem set

• Preprocess images with fillholes and watershed• Input image resize to 512x512• Predict masks with the model• Upscale the mask images with the original size after postprocesing

(e.g. fill holes)

5 Scientific Computing Facility@MPI-CBG

Page 6: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Submission the results

• Result format

• RLE encoding: Run-length encoding3

• Check the leaderboard• Public Leader Board

• Our score is 0.419 in 2nd stage while Benoit achieve better after competition

3 https://www.kaggle.com/rakhlin/fast-run-length-encoding-python

6 Scientific Computing Facility@MPI-CBG

Page 7: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Better approaches from other teams• No 1. [ods.ai] topcoders: IoU = 0.631

• Solution description: https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741• github: https://github.com/selimsef/dsb2018_topcoders/

• Unet Nuke: IoU = 0.545• Solution description: https://www.kaggle.com/c/data-science-bowl-2018/discussion/54742• github: https://github.com/nicolefinnie/kaggle-dsb2018

• Image Data sets to train• Broad Bioimage Benchmark Collection https://data.broadinstitute.org/bbbc/image_sets.html• Image Data Resource https://idr.openmicroscopy.org/about/index.html

7 Scientific Computing Facility@MPI-CBG

Page 8: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Future plans discussed• Web site for pre-test nuclei detection before consulting to Bio-

image informatics• User can submit training data with annotation• Improve the model and make the better prediction periodically• Easily change the model based on the problem context

• Flywing, neuron, etc

8 Scientific Computing Facility@MPI-CBG

Page 9: Report of Kaggle Competition - indico.mpi-cbg.de · Report of Kaggle Competition (GaHoBe-Gayathri, HongKee, Benoit) Scientific Computing Facility HongKee Moon (moon@mpi-cbg.de) 5th

Demo !9 Scientific Computing Facility@MPI-CBG