Top position in Medical Segmentation Decathlon challenge – University of Copenhagen

SCIENCE AI Centre > News > Top position in Medica...

26 November 2018

Top position in Medical Segmentation Decathlon challenge

Top ranking

Researchers from the Machine Learning Section at Department of Computer Science (DIKU) and industrial partner Cerebriu obtained a top position in the recent 2018 Medical Segmentation Decathlon. The team, named CerebriuDIKU, consisted of master’s thesis student Mathias Perslev under supervision of Akshay Pai (Cerebriu & DIKU), Erik Dam, Sune Darkner and Christian Igel (DIKU).

The participating teams were tasked to develop a system that could learn to solve 10 significantly different medical image segmentation tasks in a fully autonomous setup. The system could not utilize any task-specific information and human observers were not allowed to interact with the system at any time during training or prediction.

Team CerebriuDIKU discovered that a single, fixed topology 2D fully convolutional neural network could be applied without any hyperparameter search to the different tasks when trained under a “multi-view” scenario. Specifically, the team developed a model that simultaneously accepts images sampled from multiple views onto a 3D medical volume, facilitating a strong representation of the 3D volume while using efficient computations on 2D views. The trained model learns to observe the target as seen from multiple angles. This feature was further utilized by predicting multiple times on each subject to increase performance.

When trained in this manner, the model displayed an almost non-existing degree of overfitting and could be applied with both a fixed topology and fixed hyperparameter set to all the 10 segmentation problems, thus facilitating a simple autonomous implementation.

Team CerebriuDIKU and the remaining top-5 teams presented their methods at the 2018 MICCAI conference in Granada, Spain.

Read about the Medical Segmentation Decathlon here.

Phase 1 results: https://decathlon.grand-challenge.org/evaluation/results/

Method paper: https://1drv.ms/b/s!AhgI3jIn2dNrhcp8yLYv0_EsC97u9A

Figure 1

Figure 2