Detekcija anomalij na medicinskih slikah: od nenadzorovanega do nadzorovanega učenja

Anomaly detection in medical imaging plays a crucial role in identifying atypical anatomical patterns, supporting early diagnosis, and treatment planning. We introduce PatchCore++, a method that uses dual memory banks to store normal and anomalous features separately. This setup enables effective detection even with limited anomalous training data. PatchCore++ can utilize anomalous samples in both weakly supervised settings, where only image-level labels are available, and fully supervised settings, where both image- and pixel-level labels are provided. Extensive tests on brain MRI and liver CT datasets show that PatchCore++ outperforms existing methods when trained with fewer anomalous samples. It achieves competitive results with SuperSimpleNet, trained in both unsupervised and fully supervised settings, as well as with SegDecNet, trained in fully supervised mode. Moreover, PatchCore++ outperforms its baseline, the unsupervised PatchCore. Unlike PatchCore, which cannot utilize anomalous training samples, PatchCore++ effectively leverages them when present. We also introduce SupervisedAnomalyNet, a fully supervised method for anomaly detection and localization. We also investigate how training with different anomalous sample sizes affects performance.