Fig. 3.

Performance assessment through the area under the receiver operating curve (AUC) for the internal (COPDGene) and external (COSYCONET) test sets, for four different input configurations [0% and 20% patch-overlapping applied to Insp CT (1 channel) and Insp + ExpR (2 channels)], for the anomaly detection method (cOOpD), and for four supervised deep learning methods: end-to-end Patch Classifier with a recurrent neural network [PatClass + RNN], a multiple instance learning [MIL] with RNN as aggregation [MIL + RNN], an attention-based MIL [MIL + Att], and one representation-based [ReContrastive]