Top panel: confusion matrix showing the performance of SAD-CNN V-3 on a random sample of 200 halos from the Illustris TNG50 simulation. Each halo image has a randomly assigned surface brightness distribution and orientation. The numbers in parenthesis indicate the absolute count of each classification result. Lower panel: density map of false cases (FP+FN) as a function of the V-3 output value and the surface brightness limit (${\mu }_{lim}^r$). The red dashed line marks the decision threshold at 0.5. The total number of false cases is noted in red (FC_tot = 23). The density is normalised to the total number of false cases.