Introduction: Hepatocellular Carcinoma (HCC) diagnosis and recurrence detection rely primarily on contrast-enhanced CT (CECT) scans, which require expert radiological interpretation to characterize heterogeneous liver lesions within diverse hepatic parenchymal textures. Several benign and malignant nodules may mimic HCC across arterial or portal-venous phases, making differential diagnosis challenging. Moreover, CECT scans embed a vast amount of digital information, including radiomic data, beyond human visual perception. Artificial intelligence (AI)-driven liver and tumor segmentation, and radiomic feature extraction for histopathological characterization (tumor grading, satellitosis, and microvascular invasion, [MVI]), represents a promising strategy to assist radiological interpretation and surgical management. Aims: To create a novel dataset (HCC-ARSeg) specifically focused on untreated HCC, for the development of an AI-based framework for segmentation, diagnosis, and histological characterization from CECT scans. The study also aimed to train and subsequently identify the best-performing deep-learning (DL) architecture and supervised Machine Learning (ML) algorithms for segmentation, classification, and radiomic prediction of HCC pathological characteristics. Methods: Between 2021 and 2024, 240 patients (110 with HCC and 130 with non-HCC lesions, including angioma, neuroendocrine tumor, focal nodular hyperplasia, intrahepatic cholangiocarcinoma, and colorectal liver metastasis) were included in the HCC-ARSeg dataset. All three CECT scan phases were fully annotated by two expert radiologists. Multiple DL architectures, nnU-Net, U-Mamba, and SwinUNETR, were evaluated to identify the best-performing model for liver and HCC segmentation on both publicly available (LiTS, HCC-TACE-Seg, and WAW-TACE) and newly developed HCC-ARSeg datasets, focusing on the portal-venous phase. After preprocessing liver and tumor segmentation masks, radiomic features were extracted using PyRadiomics across all three CE phases. The resulting triple-phase 3D segmentation masks were then integrated with clinical variables (age, sex, AFP level, nodules’ number, and relative tumor volume), and the most informative radiomic features were selected. Finally, the extracted features were used to train and evaluate supervised ML algorithms (ElasticNetCV, Support Vector Machine, LightGBM, Random Forest, and XGBoost) for predicting tumor grading, satellitosis and MVI. Results: nnU-Net consistently outperformed alternative DL architectures in liver and HCC segmentation, achieving a Dice Score of 0.712 and superior boundary precision (SASD 3.903). When trained including non-HCC cases, liver segmentation performance improved (Dice`0.995), but HCC segmentation and classification accuracy decreased (Dice 0.356 and 0.533, respectively), likely due to similar CE behaviors between HCC and non-HCC lesions in portal-venous phase. In the radiomic study, ElasticNetCV achieved the best predictive performance for tumor grading (AUC 0.648), LightGBM for MVI (AUC 0.784), and Support Vector Machine for satellitosis (AUC 0.716). Predictive performance improved after inclusion of clinical variables. Conclusions: Our HCC-ARSeg dataset represents one of the most comprehensive, fully annotated, three phases CECT data resource integrated with clinical and histopathological data for AI model development in untreated HCC. The nnU-Net architecture demonstrated the highest segmentation accuracy, while radiomic analysis across all phases provided a comprehensive multiparametric framework for predicting relevant pathological features. However, the limited number of G1, MVI, and satellitosis cases hampered model learning and generalization, emphasizing the need for larger multicenter datasets.
Introduzione: Diagnosi e riscontro di recidiva di epatocarcinoma (HCC) si basano principalmente su immagini TC con mezzo di contrasto (CECT), che richiedono la valutazione del radiologo esperto, infatti l’HCC tipicamente insorge in parenchimi epatici strutturalmente alterati e sono numerose le lesioni, sia benigne che maligne, che possono mimare il comportamento contrastografico dell’HCC, rendendo complessa la diagnosi differenziale. Le immagini TC contengono numerose informazioni digitali, comprese quelle radiomiche, non percepibili all’occhio umano. Pertanto, la segmentazione automatica di fegato e tumore guidata da modelli di intelligenza artificiale (AI), associata all’estrazione di feature radiomiche per la caratterizzazione istopatologica (grading tumorale, satellitosi e invasione microvascolare [MVI]), costituiscono strumenti a supporto di interpretazione radiologica e pianificazione chirurgica. Obiettivi: Creare un database (HCC-ARSeg) esclusivamente dedicato ai casi di HCC non trattati e finalizzato allo sviluppo di modelli di AI per la segmentazione, la diagnosi e la caratterizzazione istologica dell’HCC a partire da immagini TC. Ulteriori obiettivi sono addestramento e successiva identificazione delle architetture di deep learning (DL) e degli algoritmi di machine learning (ML) con le migliori prestazioni di segmentazione, classificazione e previsione radiomica delle caratteristiche istologiche dell’HCC. Metodi: Tra il 2021 e il 2024 sono stati inclusi nel database HCC-ARSeg 240 pazienti (110 HCC e 130 non-HCC, tra cui angiomi, tumori neuroendocrini, iperplasie nodulari focali, colangiocarcinomi intraepatici e metastasi epatiche da carcinoma colorettale). Due radiologi esperti hanno annotato tutte e tre le fasi contrastografiche delle TAC incluse. Sono state valutate, in fase portale, le performance di segmentazione di fegato ed HCC, e di classificazione di HCC vs. non-HCC di diverse architetture di DL, nnU-Net, U-Mamba e SwinUNETR, per identificare il modello con le migliori prestazioni, utilizzando sia database pubblici (LiTS, HCC-TACE-Seg e WAW-TACE) che HCC-ARSeg. Successivamente, mediante PyRadiomics su tutte e tre le fasi contrastografiche sono state estratte le feature radiomiche. Le maschere 3D multifase ottenute sono state associate a variabili cliniche (età, sesso, AFP, numero di noduli e volume tumorale relativo) e sono state quindi selezionate le feature radiomiche più informative. Infine, tali feature sono state utilizzate per addestrare e valutare le performance di previsione di grading, satellitosi e MVI degli algoritmi supervisionati di ML: ElasticNetCV, Support Vector Machine, LightGBM, Random Forest e XGBoost. Risultati: L’architettura nnU-Net ha dimostrato le performance di segmentazione migliori (Dice Score di 0,712) con accurata definizione dei margini (SASD 3,903). Quando il training è stato esteso ai casi non-HCC, la performance di segmentazione del fegato e’ migliorata (Dice 0,995), ma quelle di segmentazione e classificazione dell’HCC sono peggiorate (Dice 0,356 e 0,533), spiegabile con i simili comportamenti contrastografici di HCC e non-HCC in fase portale. ElasticNetCV ha dimostrato la migliore performance predittiva per il grading tumorale (AUC 0,648), LightGBM per la MVI (AUC 0,784) e Support Vector Machine per la satellitosi (AUC 0,716). Le performance predittive sono migliorate dopo l’inclusione delle variabili cliniche. Conclusioni: HCC-ARSeg e’ uno dei database più completi e integrati con informazioni cliniche e istopatologiche per lo sviluppo di modelli di AI dedicati all’HCC. L’architettura nnU-Net ha dimostrato la massima accuratezza nella segmentazione e l’analisi radiomica ha fornito modelli di ML multiparametrici per la previsione delle principali caratteristiche istologiche.
Sviluppo di un modello multimodale di intelligenza artificiale per la segmentazione, la diagnosi e l’estrazione di caratteristiche radiomiche finalizzate alla caratterizzazione istopatologica dell’epatocarcinoma da immagini di tomografia computerizzata
ESPOSITO, GIUSEPPE
2026
Abstract
Introduction: Hepatocellular Carcinoma (HCC) diagnosis and recurrence detection rely primarily on contrast-enhanced CT (CECT) scans, which require expert radiological interpretation to characterize heterogeneous liver lesions within diverse hepatic parenchymal textures. Several benign and malignant nodules may mimic HCC across arterial or portal-venous phases, making differential diagnosis challenging. Moreover, CECT scans embed a vast amount of digital information, including radiomic data, beyond human visual perception. Artificial intelligence (AI)-driven liver and tumor segmentation, and radiomic feature extraction for histopathological characterization (tumor grading, satellitosis, and microvascular invasion, [MVI]), represents a promising strategy to assist radiological interpretation and surgical management. Aims: To create a novel dataset (HCC-ARSeg) specifically focused on untreated HCC, for the development of an AI-based framework for segmentation, diagnosis, and histological characterization from CECT scans. The study also aimed to train and subsequently identify the best-performing deep-learning (DL) architecture and supervised Machine Learning (ML) algorithms for segmentation, classification, and radiomic prediction of HCC pathological characteristics. Methods: Between 2021 and 2024, 240 patients (110 with HCC and 130 with non-HCC lesions, including angioma, neuroendocrine tumor, focal nodular hyperplasia, intrahepatic cholangiocarcinoma, and colorectal liver metastasis) were included in the HCC-ARSeg dataset. All three CECT scan phases were fully annotated by two expert radiologists. Multiple DL architectures, nnU-Net, U-Mamba, and SwinUNETR, were evaluated to identify the best-performing model for liver and HCC segmentation on both publicly available (LiTS, HCC-TACE-Seg, and WAW-TACE) and newly developed HCC-ARSeg datasets, focusing on the portal-venous phase. After preprocessing liver and tumor segmentation masks, radiomic features were extracted using PyRadiomics across all three CE phases. The resulting triple-phase 3D segmentation masks were then integrated with clinical variables (age, sex, AFP level, nodules’ number, and relative tumor volume), and the most informative radiomic features were selected. Finally, the extracted features were used to train and evaluate supervised ML algorithms (ElasticNetCV, Support Vector Machine, LightGBM, Random Forest, and XGBoost) for predicting tumor grading, satellitosis and MVI. Results: nnU-Net consistently outperformed alternative DL architectures in liver and HCC segmentation, achieving a Dice Score of 0.712 and superior boundary precision (SASD 3.903). When trained including non-HCC cases, liver segmentation performance improved (Dice`0.995), but HCC segmentation and classification accuracy decreased (Dice 0.356 and 0.533, respectively), likely due to similar CE behaviors between HCC and non-HCC lesions in portal-venous phase. In the radiomic study, ElasticNetCV achieved the best predictive performance for tumor grading (AUC 0.648), LightGBM for MVI (AUC 0.784), and Support Vector Machine for satellitosis (AUC 0.716). Predictive performance improved after inclusion of clinical variables. Conclusions: Our HCC-ARSeg dataset represents one of the most comprehensive, fully annotated, three phases CECT data resource integrated with clinical and histopathological data for AI model development in untreated HCC. The nnU-Net architecture demonstrated the highest segmentation accuracy, while radiomic analysis across all phases provided a comprehensive multiparametric framework for predicting relevant pathological features. However, the limited number of G1, MVI, and satellitosis cases hampered model learning and generalization, emphasizing the need for larger multicenter datasets.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD Thesis - Dr.pdf
embargo fino al 26/01/2029
Licenza:
Tutti i diritti riservati
Dimensione
4.62 MB
Formato
Adobe PDF
|
4.62 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/356600
URN:NBN:IT:UNIMORE-356600