Presentation of submitted article ‘Investigating Demographic Bias in Brain MRI Segmentation: A Comparative Study of Deep-Learning and Non-Deep-Learning Methods’

Submitted article

Conclusion:

• Results of UNesT and ANTs showed race matching improves segmentation accuracy
• nnU-net the only model that its performance is indifferent to the race-matching and sex-matching of training set and test set
• Sex differences observed with manual segmentation on the volumes can also be observed with biased models, whereas the race differences disappear in all but one model
• Most models show a lower overall Dice coefficient score and ESSP when trained on datasets from black demographic groups than those trained on white demographic groups.