Medical Image Anomaly Detection: From Unsupervised to Supervised Learning

Jovanoski, Hristijan (author), Skočaj, Danijel (mentor)

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.

• Language: English
• Keywords: medical imaging, anomaly detection, unsupervised learning, weakly supervised learning, fully supervised learning, PatchCore, PatchCore++.
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-172544
• COBISS.SI-ID: 248889603
• Publication date in RUL: 08.09.2025

Secondary language

• Language: Slovenian
• Title: Detekcija anomalij na medicinskih slikah: od nenadzorovanega do nadzorovanega učenja

Abstract

Detekcija anomalij na medicinskih slikah ima ključno vlogo pri prepoznavanju neobičajnih anatomskih vzorcev, podpira zgodnjo diagnostiko in načrtovanje zdravljenja. V diplomskem delu predstavimo metodo PatchCore++, ki uporablja dvojne pomnilniške množice za ločeno shranjevanje značilk normalnih in anomalnih vzorcev, hkrati pa omogoča tudi učinkovito detekcijo ob omejenem naboru anomalnih vzorcev v učni množici. PatchCore++ lahko deluje v šibko nadzorovanih pogojih, kjer so oznake na voljo le na nivoju slik, ter v popolnoma nadzorovanih pogojih, kjer so oznake podane tako na ravni slik kot tudi pikslov. Rezultati poskusov na slikah možganov (MR) in jeter (CT) kažejo, da PatchCore++ presega obstoječe metode pri manjšem številu anomalnih vzorcev. Metoda dosega konkurenčne rezultate s SuperSimpleNet, ki deluje v nenadzorovanem in popolnoma nadzorovanem načinu, ter s SegDecNet, ki deluje le v popolnoma nadzorovanem načinu. Poleg tega PatchCore++ daje boljše rezultate od svoje osnovne različice PatchCore, saj učinkovito izkorišča tudi anomalne podatke. Poleg tega predstavimo še SupervisedAnomalyNet, popolnoma nadzorovano metodo za detekcijo in lokalizacijo anomalij, ter preučimo vpliv učenja z različnim številom anomalnih vzorcev na uspešnost.

• Keywords: medicinske slike, detekcija anomalij, nenadzorovano učenje, šibko~nadzorovano~učenje, popolnoma~nadzorovano~učenje, PatchCore, PatchCore++.