This repository provides code snippets used to generate 11th place solution in Prostate cANcer graDe Assessment (PANDA) challenge Additional description is provided in this discussion , solution writeup and public kernels demonstrating generation of tiles , training a model , and generation of predictions
The files in the repository demonstrate generation of tiles (PANDA_train.ipynb), generation of tiles with advanced tile selection (PANDA_create_adv_tiles.ipynb), pretraining the models on low resolution tiles (PANDA_pretrain.ipynb), training the final models (PANDA_train.ipynb), and training models with advanced tile selection (PANDA_train_adv_tiles.ipynb), which is not used to generate the final submission in the competition.
The performance of the basic setup (averaged over 5 4-fold models trained with different data splits) is 0.932+-0.002 at private LB and 0.910+-0.004 at public LB, with performance of the best model of 0.938/0.916. The advanced tile selection, based on the algorithm posted in this kernel, gives 0.935+-0.006 and 0.910+-0.003 at private and public LB, respectively, with the performance of the best model of 0.941/0.913.
The final models are available at this link and the final submission is generated by this kernel based on majority voting ensemble. An inference kernel illustrating use of the advanced tile selection (0.942/0.911 private/public LB) is provided by this link, and a kernel generating a prediction based on a simple average over predictions of all models (0.934/0.914 private/public LB) is provided by this link .