In hepato-pancreato-biliary surgery, success depends on accurately understanding patient-specific anatomy and clinical context. Yet, high anatomical variability often makes standard imaging insufficient. Generating reliable 3D reconstructions from CT scans thus offers a concrete way to enhance anatomical insight and support clinical reasoning. This work introduces task-specific AI solutions for 3D reconstructions developed and validated through a rigorous, clinically guided methodology. The first contribution is a systematic review of 130 studies on pancreatic segmentation, highlighting recurring limitations in existing models, such as insufficient subregional analysis and lack of large-scale validation. Building on these insights, a 2.5D AI model was trained on over 1,300 CT scans annotated by AIMS Academy. The proposed approach outperformed the state of the art on the public AMOS dataset, achieving a DSC of 0.87 (HD = 12.73 mm, ASSD = 0.66 mm) for pancreas segmentation. The work also provided quantitative and qualitative insights into the data collection and annotation effort, highlighting the hidden cost of building high-quality clinical datasets. The second contribution focuses on automatic segmentation of hepatic vasculature, using a dataset of 385 subjects with distinct labeling of portal and hepatic branches. The D2-RD-UNet model integrates data-driven preprocessing, vessel enhancement filters, automatic topological corrections, and diameter-based metrics, outperforming state-of-the-art approaches even on external benchmarks. Overall, the proposed model outperformed the state of the art on the public 3D-IRCADb-01 dataset, achieving a DSC of 0.62 (HD95 = 15.63 mm, ASSD = 2.90 mm) for hepatic vessels and 0.68 (HD95 = 16.72 mm, ASSD = 2.57 mm) for portal vessels. Finally, the work includes tools to support transferability, such as a web app for anatomical exploration, a study on vascular datasets, and 3D rendering solutions. Overall, the thesis supports the transition from a protocol-based surgical approach to a more in-depth and personalized anatomical understanding, transforming 3D digital reconstructions into a concrete, navigable, and clinically valuable resource.
Nella chirurgia epato-pancreato-biliare, il successo dell’intervento dipende da un’accurata comprensione dell’anatomia del paziente. L’elevata variabilità anatomica spesso limita le informazioni ottenibili dalle sole immagini mediche. Le ricostruzioni 3D delle strutture di interesse a partire da TAC rappresentano un mezzo concreto per migliorare la conoscenza anatomica e supportare il processo decisionale clinico. Il presente lavoro introduce soluzioni AI task-specifiche per ricostruzione 3D, sviluppate e validate secondo una metodologia scientifica guidata da bisogni clinici reali. Il primo contributo è una revisione sistematica di 130 studi sulla segmentazione pancreatica, che ha evidenziato limiti ricorrenti nei modelli esistenti, come la scarsa attenzione alla variabilità subregionale e la mancanza di validazioni su larga scala. Da queste basi nasce un modello AI 2.5D, testato su oltre 1.300 TC annotate da AIMS Academy. L’approccio proposto ha superato lo stato dell’arte sul dataset pubblico AMOS, raggiungendo un DSC di 0,87 (HD = 12,73 mm, ASSD = 0,66 mm) per la segmentazione del parenchima pancreatico. Il lavoro fornisce inoltre indicazioni quantitative e qualitative sull’effort necessario per la raccolta e annotazione dei dati, evidenziando il costo implicito della generazione di dataset clinici di qualità. Il secondo contributo riguarda la segmentazione automatica della vascolarizzazione epatica, su un dataset di 385 soggetti con etichettatura distinta dei rami portali ed epatici. Il modello D2-RD-UNet integra un preprocessamento data-driven, filtri per l'enhancement vascolare, correzioni topologiche automatiche e metriche localizzate basate sul calibro. Nel complesso, il modello proposto ha superato lo stato dell’arte sul dataset pubblico 3D-IRCADb-01, ottenendo un DSC di 0,62 (HD95 = 15,63 mm, ASSD = 2,90 mm) per le vene epatiche e 0,68 (HD95 = 16,72 mm, ASSD = 2,57 mm) per le vene portali. Infine, il lavoro propone strumenti per favorire la trasferibilità, tra cui l'integrazione in una web app, un’analisi dei dataset e soluzioni di rendering 3D. Nel complesso, la tesi incentiva il passaggio ad una comprensione anatomica più profonda e personalizzata, rendendo le ricostruzioni 3D uno strumento concreto e clinicamente utile.
From raw data to clinical strategy: AI-based digital twins for patient specific vascular and parenchymal modeling in HPB surgery
Cavicchioli, Matteo
2025/2026
Abstract
In hepato-pancreato-biliary surgery, success depends on accurately understanding patient-specific anatomy and clinical context. Yet, high anatomical variability often makes standard imaging insufficient. Generating reliable 3D reconstructions from CT scans thus offers a concrete way to enhance anatomical insight and support clinical reasoning. This work introduces task-specific AI solutions for 3D reconstructions developed and validated through a rigorous, clinically guided methodology. The first contribution is a systematic review of 130 studies on pancreatic segmentation, highlighting recurring limitations in existing models, such as insufficient subregional analysis and lack of large-scale validation. Building on these insights, a 2.5D AI model was trained on over 1,300 CT scans annotated by AIMS Academy. The proposed approach outperformed the state of the art on the public AMOS dataset, achieving a DSC of 0.87 (HD = 12.73 mm, ASSD = 0.66 mm) for pancreas segmentation. The work also provided quantitative and qualitative insights into the data collection and annotation effort, highlighting the hidden cost of building high-quality clinical datasets. The second contribution focuses on automatic segmentation of hepatic vasculature, using a dataset of 385 subjects with distinct labeling of portal and hepatic branches. The D2-RD-UNet model integrates data-driven preprocessing, vessel enhancement filters, automatic topological corrections, and diameter-based metrics, outperforming state-of-the-art approaches even on external benchmarks. Overall, the proposed model outperformed the state of the art on the public 3D-IRCADb-01 dataset, achieving a DSC of 0.62 (HD95 = 15.63 mm, ASSD = 2.90 mm) for hepatic vessels and 0.68 (HD95 = 16.72 mm, ASSD = 2.57 mm) for portal vessels. Finally, the work includes tools to support transferability, such as a web app for anatomical exploration, a study on vascular datasets, and 3D rendering solutions. Overall, the thesis supports the transition from a protocol-based surgical approach to a more in-depth and personalized anatomical understanding, transforming 3D digital reconstructions into a concrete, navigable, and clinically valuable resource.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_11_Cavicchioli.pdf
accessibile in internet per tutti
Descrizione: Tesi Dottorato
Dimensione
50.54 MB
Formato
Adobe PDF
|
50.54 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/245258