Explainable artificial intelligence for disease diagnosis from images

Kogovšek, Žan (author), Emeršič, Žiga (mentor)

Accurate and timely diagnosis of diseases from medical images is a critical challenge in healthcare, often limited by the availability of expert clinicians and the complexity of image interpretation. This thesis addresses the problem of automated human disease diagnosis from medical images by developing and evaluating state-of-the-art deep learning models, as well as integrating explainable artificial intelligence (XAI) techniques. We systematically benchmarked a range of advanced architectures across multiple datasets and disease domains, including skin cancer, diabetic retinopathy, and pneumonia. Our approach emphasizes not only high diagnostic accuracy but also model transparency, with a particular focus on the quantitative evaluation of XAI methods such as CAM-based visualizations and example-based retrieval. The primary focus of this work and best results were achieved on the ISIC 2018 skin lesion dataset, where our methods reached a balanced multiclass accuracy of 85.1%, placing us among the top performers on the official leaderboard.

• Language: English
• Keywords: Medical image analysis, Deep learning, Class imbalance, Explainable artificial intelligence
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-171825
• COBISS.SI-ID: 248404227
• Publication date in RUL: 03.09.2025

Secondary language

• Language: Slovenian
• Title: Razložljiva umetna inteligenca za diagnozo bolezni na osnovi slik

Abstract

Pravočasna in natančna diagnoza bolezni na podlagi medicinskih slik predstavlja enega izmed osrednjih izzivov sodobne medicine. Pomanjkanje zdravstvenih specialistov in zahtevnost interpretacije slikovnih podatkov dodatno otežujeta zanesljivo postavljanje diagnoz. Magistrsko delo naslavlja ta izziv z razvojem naprednih modelov umetne inteligence za avtomatizirano diagnozo bolezni na osnovi medicinskih slik, pri čemer je poseben poudarek namenjen integraciji metod razložljive umetne inteligence. Izvedena je bila sistematična analiza različnih modelov na več podatkovnih množicah in za različne bolezni, vključno s kožnim rakom, diabetično retinopatijo in pljučnico. Poleg visoke diagnostične natančnosti se delo osredotoča na zagotavljanje transparentnosti modelov z uporabo vizualnih razlag in razlag na podlagi primerov, ki so bili tudi kvantitativno ovrednoteni. Najboljši rezultat smo dosegli na podatkovni množici ISIC 2018 za prepoznavo kožnega raka, kjer je bila dosežena uravnotežena večrazredna natančnost 85,1%, kar uvršča razviti pristop med vodilne na uradnih lestvicah.

• Keywords: Analiza medicinskih slik, Globoko učenje, Neravnovesje razredov, Razložljiva umetna inteligenca