Figure 3.

Probabilistic model. Our modeling assumption is that each prediction of radiologists and DNNs is influenced by four latent variables. ${\widehat{y}}_{r,s}^{(n)}$ is radiologist (DNN) r’s prediction on case n filtered with severity s. As for the latent variables, $b_g$ represents the bias for subgroup g, ${\mu }^{(n)}$ is the bias for case n, ${\gamma }_{s,g}$ is the effect that low-pass filtering with severity s has on lesions in subgroup g, and ${\nu }_{r,g}$ is the idiosyncrasy of radiologist (DNN) r on lesions in subgroup g. Our analysis relies on the posterior distribution of ${\gamma }_{s,g}$, as well as the posterior predictive distribution of ${\widehat{y}}_{r,s}^{(n)}$. The other latent variables factor out potential confounding effects. Figure created with drawio v13.9.0 https://github.com/jgraph/drawio.