Deep learning in the magnetic resonance imaging-based diagnostics of focal liver lesions using hepatocyte-specific contrast agents
Stollmayer Róbert
KÁROLY RÁCZ CONSERVATIVE MEDICINE PROGRAM
Dr. Fekete Andrea
SE Transzplantációs és Sebészeti Klinika
2025-10-27 13:00:00
Gastroenterology
Dr. Molnár Béla
Dr. Kaposi Novák Pál
Dr. Földes-Lénárd Zsuzsanna
Dr. Horváth András
Dr. Szijártó Attila
Dr. Lotz Gábor
Dr. Géher Pál
The current thesis discusses the development and evaluation of recently
introduced deep learning (DL) methods for the automated classification and description
of focal liver lesions (FLLs) based on magnetic resonance imaging (MRI) using
hepatocyte-specific contrast agents (HSCs). As the first study of its kind in Hungarian
literature, it demonstrates that convolutional neural networks (CNNs) can effectively
differentiate focal nodular hyperplasia (FNH), hepatocellular carcinoma (HCC), and liver
metastases (MET) based on both two-dimensional (2D) and three-dimensional (3D)
information. Furthermore, various radiological features of focal liver lesions are
automatically identifiable by CNNs, such as DenseNets and EfficientNets. These
advancements enable the development of downstream imaging methods, particularly in
deep learning reconstruction, which requires the definition of a ground truth/reference to
which a generated image is compared.
The study methodology involved the manual collection and preparation of large
quantities of MRI data. Data processing includes anonymization, resampling, and
alignment of MRI scans, followed by the creation of training, validation, and test datasets.
All architectures were employed within the MONAI framework for analysis and custom
written computer code was written by our research group for both analyses.
Our first study found that both 2D and 3D CNNs effectively differentiate between
FNH, HCC, and MET with AUCs above 0.90. In our second study EfficientNetB0 was
identified as the top-performing model for radiological feature identification, achieving
the highest validation mean AUC (0.9147) after 480 epochs.
The integration of such models into clinical practice faces challenges, such as the
need for larger, multi-institutional datasets and further validation studies as well as more
detailed large scale data annotation.
Our research concludes that DL techniques, particularly CNNs, are promising
tools for enhancing MRI-based diagnosis of FLLs. The use of HSC-enhanced MRI
combined with advanced DL models shows high diagnostic accuracy, aiding early and
precise diagnosis of liver conditions. In summary, this dissertation demonstrates
significant advancements in applying DL to medical imaging, providing a robust
framework for future research and clinical integration in diagnosing focal liver lesions.
