Code used for competition submission. This solution got me in the top 14% on the private leaderboard (out of 1157 competitors). Unfortunately there was a data leak near the end of the competition. By exploiting the leak, some participants were able to get a perfect score.

The goal is to detect the presence of metastatic cancer in image patches taken from pathology scans (i.e. binary classification task).

I tried different CNN architectures and trained with weights initialized based on ImageNet. The main steps leading to performance gain were:

• data augmentation (using the imgaug library)
• test time augmentation
• ensembling

Training took about 4-6 hours (depending on CNN architecture) on a P4000 GPU for 12 epochs. Increasing the number of epochs might lead to performance improvements (although I saw signs of overfitting after about 9-10 epochs already).

The final ensemble yielded a 0.9696 AUC on the private leaderboard.

• tensorflow 1.5.0
• keras 2.2.4
• opencv-python
• imgaug
• sklearn
• pandas