UBC Ocean Kaggle competition – Meet the winners

In the data-driven world we live in, Kaggle competitions have become a popular forum for solving complex problems, particularly in precision medicine and healthcare. Solutions developed in previous competitions have contributed to improve the detection of disease using medical images in several therapeutic areas, such as diabetic retinopathy, lung cancer, and melanoma. (reference)

Owkin data scientists Etienne Andrier, Valentina Di Proietto, and Jean-Baptiste Schiratti, teamed up with former Owkin employee Simon Jegou to participate in a Kaggle competition hosted by the University of British Columbia (UBC) and aimed at improving classification of the five common ovarian cancer subtypes on histopathology images and to detect outliers (rare subtypes).

The competition, named UBC-OCEAN (UBC Ovarian Cancer Subtype Classification and Outlier Detection), saw an impressive 1,300 teams participating. When the Owkin team looked at the leaderboard, they found themselves at the top of the ranking.

The solution they proposed is an ensemble of Chowder models trained on top of Phikon, a ViT-B model pretrained on TCGA using iBOT. The team then used high entropy predictions to detect outliers. The ensembling was performed over 15 different repetitions and folds, and each was initialized 50 times, bringing the total number of models to 750, the outputs of each fold and repetition were calibrated using the validation sets, which lead to excellent generalization. Even better performance was obtained on the private dataset than on the public dataset. Both Phikon and Chowder have been previously developed by Owkin data scientists. Chowder is a multiple-instance learning algorithm, and Phikon is a cutting-edge histology foundation model (FM). Phikon is publicly available on GitHub and HuggingFace. The team ranking 6th also proposed a solution involving Chowder and Phikon. This not only demonstrates Owkin’s contribution to the development of quality models to tackle real-world medical problems, but also shows that the AI/ML expertise of Owkin’s team played a decisive role.

One of the key takeaways from this achievement is once again the efficiency of specialized FMs for medical imaging, and more broadly the potential of FMs for answering medical questions. This is only scratching the surface of what AI can do in healthcare and precision medicine, using a single modality. We’re excited by the prospect of broadening machine capabilities to include the wealth of medical data that healthcare is generating, from electronic health records, to medical imaging and multi omics data. Developing the kind of multimodal AI that can connect all these scales of biology will allow us to actualize the potential of AI in medicine.

At Owkin we are committed to using AI to bridge the gap in understanding complex medical biology. This achievement is a testament to our unwavering dedication to excellence and innovation in the intersection of artificial intelligence and healthcare.