Kaggle Prostate cANcer graDe Assessment (PANDA) Challenge

Here is the solution code of Team ChienYiChi for https://www.kaggle.com/c/prostate-cancer-grade-assessment/overview.

• kaggle profile: https://www.kaggle.com/ericji
• kaggle solution discussion: https://www.kaggle.com/c/prostate-cancer-grade-assessment/discussion/169114

Dependencies

pip install -r requirements.txt

Experiments Record

tile images model

type	model	private kappa	public kappa	local kappa	karolinska kappa	radboud kappa	fold num	image size	num tiles	epoch	TTA
cls	tiles-resnext50-netvlad	0.896	0.879	0.8602	0.8884	0.8089	0	256	20	27	8
cls	tiles-eb0-netvlad	0.882	0.849	0.8762	0.877	0.851	0	256	20	26	8
cls	tiles-eb0-netvlad	0.857	0.856	0.8834	0.8714	0.8692	0	256	36	22	8
cls	tiles-resnet34-netvlad	0.87	0.848	0.8745	0.8697	0.8522	0	256	20	28	8
reg	tiles-eb0-netvlad	0.899	0.859	0.8777	0.8820	0.8470	0	256	20	29	8
reg	tiles-eb0-netvlad	0.920	0.881	0.8952	0.8976	0.8704	0	256	36	28	8
reg	tiles-eb0-netvlad	0.903	0.893	0.886	0.8979	0.8464	1	256	36	22	8
reg	tiles-eb4-netvlad train with BRS（blue ratio selection）,test without BRS	0.909	0.90	0.8826	0.9047	0.8335	1	256	36	26	8
reg	tiles-eb4-netvlad test with BRS	0.913	0.896	0.8826	0.9047	0.8335	1	256	36	26	8
reg	tiles-eb0-netvlad with attention model(128 tiles) to select tiles	0.899	0.88	0.8833	0.9034	0.8367	1	256	16	27	8 TTA only for score model
reg	tiles-eb0-netvlad with attention model to select tiles	0.910	0.88	0.8833	0.9034	0.8367	1	256	16	27	8 TTA only for score model
reg	tiles-eb0-netvlad with attention model to select tiles	0.813	0.874	0.8859	0.8945	0.8481	1	256	36	25	8 TTA only for score model
reg	tiles-eb4-netvlad with attention model to select tiles	0.908	0.897	0.8766	0.9035	0.8246	1	256	16	27	8 TTA only for score model
reg	newcv tiles-eb4-netvlad with attention model to select tiles	0.904	0.899	0.8812	0.8958	0.8437	1	256	16	27	8 TTA only for score model
ord reg	newcv tiles-eb0-netvlad with attention model to select tiles	0.913	0.884	0.8958	0.8972	0.8732	1	256	16	27	8 TTA only for score model
ord reg	newcv tiles-eb4-netvlad with attention model to select tiles	0.900	0.879	0.887	0.9006	0.8524	1	256	16	27	8 TTA only for score model
reg	stitch-tiles-regnety_800m with attention model to select tiles	0.911	0.893	0.8935	0.8872	0.8757	1	256	16	28	8 TTA only for score model

How to Run

• generate tiles using preprocess.py

train one-stage model

1. set model type and hyperparameters in config.py
2. change model function in train.py

train two-stage model (attention model + score model)

1. set model type and hyperparameters in config.py
2. change the model function to the efficienet model with attention layer in train.py
3. generate tiles weights using generate_weights.py which will output a tiles weights csv file
4. set model type and hyperparameter again in config.py if you want to change the model type , for example regression or ordinal regression
5. change the model function in train.py ,for example, efficientnet with NetVlad layer

References

Thanks everyone who shared their ideas on Kaggle discussion, I learned a lot from them.

• https://www.kaggle.com/iafoss/panda-concat-tile-pooling-starter-0-79-lb
• https://www.kaggle.com/haqishen/panda-inference-w-36-tiles-256
• https://github.com/facebookresearch/pycls
• https://github.com/loadder/netVLAD-pytorch
• https://arxiv.org/abs/1511.07247