Liver MRI classification approach to improve screening in cirrhotic patients

Hepatocellular carcinoma (HCC) is the most common malignant liver tumor and represents one of the leading causes of cancer-related mortality worldwide. Early detection of HCC is crucial to improve survival outcomes in patients with chronic liver disease and imaging technique such as Magnetic Resonance is a powerful tool in HCC detection and characterization. However, the gold standard in HCC diagnosis through conventional magnetic resonance imaging rely on hepatobiliary contrast agents, raising clinical concerns regarding scan duration, procedural complexity, and patient safety. For this reason, this study investigates whether an automated radiomics-driven machine learning framework can accurately support early HCC classification and meanwhile evaluates the feasibility of contrast-agent-free screening through MRI. A retrospective cohort of 289 subjects (107 HCC, 105 hepatopathy/cirrhosis, and 82 healthy controls) underwent 1.5T multiparametric liver MRI, including pre-contrast T1-weighted, T2-weighted HASTE, diffusion-weighted imaging with ADC maps, and multiphase dynamic contrast-enhanced series (arterial, portal venous, and delayed). Two spatial localization paradigms were investigated for radiomic features extraction: an automated whole-liver segmentation using MRSegmentator and a localized slice-based slab extraction centered on the most suspicious lesion detected by a Siemens deep learning prototype (MR Auto Processor Liver 1.0). Finally, standardized 2D PyRadiomics features (IBSI-compliant) were extracted and screened through an ANOVA-based filter with collinearity pruning (|r|>0.85). The resulting biomarker signature was fed into a Random Forest classifier optimized via Bayesian search within a cross-validation scheme. In distinguishing HCC from healthy controls, the whole liver approach trained solely on pre-contrast sequences achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.772±0.020, performing comparably to the slice-based configuration (p=0.6408). In differentiating HCC from cirrhotic parenchyma, the whole liver model failed under non-contrast conditions (AUC=0.504±0.027) and remained suboptimal even post-contrast (AUC=0.668±0.033). With the slice-based approach, the classification improved to an AUC of 0.867±0.057 (p=0.0022, demonstrating its superiority over the whole-liver configuration), meanwhile, non-contrast slice-based models experienced a significant performance drop (AUC=0.780±0.092). Leave-one-sequence-out ablation study and standalone analyses confirmed that post-contrast dynamic acquisitions carry essential discriminative weight that non-contrast sequences cannot fully substitute. In conclusion, this study demonstrates that radiomic features extracted from non-contrast liver MRI are insufficient to reliably differentiate HCC from cirrhotic tissue, confirming that contrast-enhanced imaging remains indispensable for an accurate diagnosis. Conversely, in the less challenging classification task (HCC vs. healthy), the proposed machine learning pipeline achieved promising results using exclusively pre-contrast sequences. Overall, the architecture shows notable potential for HCC classification and further refinements are necessary. Future work should expand the patient cohort, incorporate advanced deep learning methodologies, and implement accurate volumetric lesion segmentation to improve diagnostic capabilities.

Il presente studio indaga l'efficacia di una pipeline radiomica automatizzata di machine learning nella diagnosi precoce del carcinoma epatocellulare (HCC), valutando la fattibilità di un protocollo di screening tramite risonanza magnetica privo di mezzo di contrasto. Su una coorte retrospettiva di 289 soggetti (107 HCC, 105 cirrotici, 82 sani) sottoposti a RM multiparametrica a 1.5T, sono stati confrontati due approcci di estrazione feature: una segmentazione automatica dell'intero fegato con MRSegmentator e un'estrazione localizzata sulla lesione (slice-based). Le feature selezionate, conformi a IBSI, sono state processate mediante Random Forest ottimizzato via ricerca Bayesiana per classificare i soggetti nei due diversi approcci. Nella discriminazione tra HCC e controlli sani, il modello pre-contrasto su intero fegato ha ottenuto un'AUC di 0.772±0.020, paragonabile alla configurazione slice-based (p=0.6408). Al contrario, nel differenziare l'HCC dal parenchima cirrotico, l'approccio sull'intero fegato è fallito senza mezzo di contrasto (AUC=0.504±0.02), mentre il modello slice-based con sequenze contrastografiche ha raggiunto un'AUC di 0.867±0.057, esibendo un calo significativo in assenza di contrasto (AUC=0.780±0.092). I risultati evidenziano che, sebbene le sequenze non contrastografiche mostrino potenziale nello screening rispetto a soggetti sani, l'imaging con mezzo di contrasto e un approccio localizzato rimangono indispensabili per distinguere accuratamente l'HCC dal tessuto cirrotico.