deepsense.ai again among the leaders in the Kaggle competition
deepsense.ai Data Scientist Jakub Pingielski has won a gold medal in Kaggle’s competition HuBMAP + HPA – Hacking the Human Body. Jakub took 12th place out of 1175 participants, coming in only 1.5 points behind the winner.
The competition was connected with the Human BioMolecular Atlas Program (HuBMAP), working to create a Human Reference Atlas at the cellular level. Sponsored by the National Institutes of Health (NIH), HuBMAP and Indiana University’s Cyberinfrastructure for Network Science Center (CNS) have partnered with institutions across the globe for this endeavor. A major partner is the Human Protein Atlas (HPA), a Swedish research program aiming to map the protein expression in human cells, tissues, and organs, funded by the Knut and Alice Wallenberg Foundation.
The task for the contestants was to identify and segment functional tissue units (FTUs) across five human organs. Participants developed models using a dataset of tissue section images, with the best submissions segmenting FTUs as accurately as possible.
“In my opinion, participating in Kaggle competitions is a great opportunity to learn state-of-the-art machine learning models and techniques from leading practitioners. I decided to participate in this contest while working on a commercial project in a similar domain at deeepsense.ai. The lessons learned from this competition enabled me to provide best-in-class performance for our client. Moreover, competitions like this are very satisfying to participate in since you know that your solution will be used for a good cause,”
said Jakub.
The key to achieving the best results was the application of very strong data augmentations while training deep neural networks. By doing so, Jakub was able to train a model that was robust to domain shift and performed equally well on microscopic images from different medical databases. The core of the training pipeline was based on the Vision Transformer model which has proven to perform better than models based on Convolutional Neural Networks. Additionally, model ensembling and test time augmentation made it possible to squeeze out model performance.
The results of the competition were of significant value – with a better idea of the relationship between cells, researchers have obtained further insight into the function of cells that impact human health. Furthermore, the Human Reference Atlas constructed by HuBMAP will be freely available for use to researchers and pharmaceutical companies alike, potentially improving and prolonging human life.
