Skip to content
This repository has been archived by the owner on Jun 22, 2022. It is now read-only.

Latest commit

 

History

History
77 lines (63 loc) · 2.97 KB

REPRODUCE.md

File metadata and controls

77 lines (63 loc) · 2.97 KB
Raw

Intro

This is an open solution to the Data Science Bowl 2018 based on the topcoders winning solution from ods.ai.

Goal

Implement winning solution described by topcoders and reproduce their results using our Tech stack, mainly steppy and steppy-toolkit.

Results (so far)

0.577 Local CV

0.457 Stage 1 LB

Solution write-up

Preprocessing

  • Overlay binary masks for each image is produced
  • Borders are produced using dilated watershed lines
  • Normalization as on ImageNet

Differences with topcoders solution:

  • Borders width doesn't depend on nuclei size

Augmentations

  • Flips u/d and l/r
  • Rotations with symmetric padding
  • piecewise affine transformation
  • perspective transform
  • inverting colors
  • contrast normalization
  • elastic transformation
  • adding random value to pixels (elementwise and uniformly, in RGB and HSV)
  • multiplying pixels by random value (elementwise and uniformly, in RGB and HSV)
  • channel shuffle
  • Gaussian, average and median blurring
  • sharpen, emboss

Differences with topcoders solution:

  • No color to gray and gray to color
  • We didn't know how often and how hard were these augmentations, if they were OneOf or SomeOf etc.

Network

  • Unet with pretrained Resnet101 or Resnet152 encoders
  • First network with softmax activation function and 3 channels: [background, masks - borders, borders] for predicting borders
  • Second network with sigmoid activation function and 2 channels: [masks, borders] for predicting full masks

Training

  • Adam optimizer
  • Initial lr 1e-4
  • Batch size of 36 (2 GPUs) or 72 (4 GPUs)
  • Training on random crops of size 256x256
  • Inference on full images padded to minimal size fitting to network (i.e. dimensions must be divisible by 64)
  • TTA (flips, rotations)

Differences with topcoders solution:

  • No info about inference in the write up, maybe it was done using sliding window not on full images.
  • Larger batchsize.

Loss function

  • 1st network: Cross Entropy with Dice (not on background)
  • 2nd network: BCE with Dice
  • Averaging Dice Loss over number of classes didn't change the results

Postprocessing

  • Different thresholds are used for masks (2nd network) for retrieving seeds and final masks
  • Seeds for watershed are calculated as masks (2nd network) - borders (1st network)
  • Small mask instances and seeds are dropped
  • Watershed using labeled seeds as markers and masks (2nd network) as masks

External data

We included data from:

Not implemented from topcoders solution

  • 2nd level model
  • model ensembling