Given a biomarker regression network, what is the effect of training with different cost functions?
In this conference publications, Sierra’s researcher investigate the influence of training with L1, L2, L1 in logarithmic scale or L2 in logarithmic scale for the problem of Agatston score regression from CT volumetric images center at the heart. There is a trade-off between biomarker accuracy and risk-group accuracy. For this particular problem, using L1 in a logarithmic scale achieves best biomarker correlation coefficient and risk accuracy group classification.
The full article can be found at: https://link.springer.com/chapter/10.1007/978-3-030-00946-5_33