Data & Augmentations

Data

Clustering

• Allen(주최측) 에서 나누어둔 이미지별 Class : https://www.kaggle.com/c/data-science-bowl-2018/discussion/48130
  • ./metadata/allencls.csv
• K-means Clustering technique has been used to split test and valid data set

External Data

Issue #36

• H&E Dataset [USED]

  • Count : 30
  • https://www.kaggle.com/voglinio/external-h-e-data-with-mask-annotations
  • https://nucleisegmentationbenchmark.weebly.com/

• TNBC Dataset [Used]

  • Count : 50
  • https://www.kaggle.com/branislav1991/converting-tnbc-external-data-to-dsb2018-format/notebook
  • /data/public/rw/datasets/dsb2018/extra_data_tnbc

• Harvard Crowd-Sourced Nuclei Annotation Data

  • Data Description:
    • The initial image color is purple and is gray-colored ("gray")
  • https://becklab.hms.harvard.edu/software/psb-crowdsourced-nuclei-annotation-data-1
  • File Directories:
    • /data/public/rw/datasets/dsb2018/origin_ext_train_kmeans
    • /data/public/rw/datasets/dsb2018/origin_ext_valid_kmeans

Score Board

Model	Data Set	Data Aug Method	Valid Loss	Valid mIOU(inference)	Score	Kaggle LB
Unet	Default	Default	0.0668	0.4710	0.4999	0.471
Unet	Default	Default2	0.0674	0.4723	0.4994	-
Unet	Default	Transparency	0.0636	0.4743	0.4992	0.472
Unet	Default	Thick line	0.0637	0.4722	0.4952	0.494
Unet	Harvard(purple_gray)	-	0.0769	0.3616	0.4228	0.461
Unet	Harvard(purple)	Thick line	0.0672	0.3854	0.5009	-
Unet	Harvard(purple)	-	0.0683	0.3842	0.4955	-
Unet	Harvard(purple)	transparent	0.0671	0.3823	0.4912	-
Unet	Harvard(purple_gray)	transp. & Thick	0.0880	0.3649	0.4230	-
Unet	Harvard(gray)	-	0.0602	0.3887	0.5024	0.487
Unet	Harvard(gray)	-	0.1355	0.3718	0.4548	-