Leveraging state-of-the-art architectures by enriching training information - a case study

Our working hypothesis is that key factors in COVID-19 imaging are the available
imaging data and their label noise and confounders, rather than network
architectures per se. Thus, we applied existing state-of-the-art convolution neural
network frameworks based on the U-Net architecture, namely nnU-Net [3],
and focused on leveraging the available training data. We did not apply any
pre-training nor modi ed the network architecture.
First, we enriched training information by generating two additional labels for
lung and body area. Lung labels were created with a public available lung segmentation
network and weak body labels were generated by thresholding. Subsequently,
we trained three di erent multi-class networks: 2-label (original background
and lesion labels), 3-label (additional lung label) and 4-label (additional
lung and body label). The 3-label obtained the best single network performance
in internal cross-validation (Dice-Score 0.756) and on the leaderboard (Dice-
Score 0.755, Haussdor 95-Score 57.5).
To improve robustness, we created a weighted ensemble of all three models, with
calibrated weights to optimise the ranking in Dice-Score. This ensemble achieved
a slight performance gain in internal cross-validation (Dice-Score 0.760). On the
validation set leaderboard, it improved our Dice-Score to 0.768 and Haussdor 95-
Score to 54.8. It ranked 3rd in phase I according to mean Dice-Score.
Adding unlabelled data from the public TCIA dataset in a student-teacher manner
signi cantly improved our internal validation score (Dice-Score of 0.770).
However, we noticed partial overlap between our additional training data (although
not human-labelled) and  nal test data and therefore submitted the
ensemble without additional data, to yield realistic assessments.