Fig. 6: Model performance generalization across images from different scanners.

a Images acquired from different scanners exhibit diverse appearances. Without applying generalization techniques, these variations lead to inconsistent AI results, as shown in (b). To mitigate the impact of appearance variations on AI model performance, images from various scanners underwent image appearance migration, reducing appearance differences. Additionally, color augmentation was employed during the generalization process, training the model using the same method as previously described. c and d display images after appearance migration and their corresponding outputs from AI model with color augmentation, respectively. With the implementation of the generalization technique, the AI outputs demonstrated increased consistency. e And the corresponding AI outputs better aligned with the ground truth annotations. f The macro average F1 score of each scanner dataset before and after the generalization techniques were applied. Notably, Gleason Pattern 5 was excluded due to the limited availability of annotated GP5 regions in cases scanned by multiple scanners.