Blood supports the exchange of oxygen, nutrients and minerals be- tween di↵erent tissues of the human body. A detailed 3D visualization of the total vascular network provides useful information about the vascular structure and organization and the distribution of the blood flow. To isolate the 3D vasculature from the background, semantic segmentation must be performed on CE-CT images, stained with a novel contrast agent (polyoxometalate) that highlight the blood ves- sels. Deep learning approaches for semantic segmentation are promis- ing and solve the problems present in the manual and semi-automated methods (time-demanding, lack of experts, prone to errors, onerous to implement in a full 3D approach). However, the absence of an existing datasets of already segmented images, prevents the training phase. To solve these issues, the master thesis project is composed by two di↵er- ent but dependent steps: first, an artificial 3D synthetic data generator is developed, to reproduce realistic 3D blood vessels network, in order not to be dependent anymore from the real data. The images are gen- erated allowing to build large datasets with synthetic images and their already segmented pairs, to be used during neural network training. Particular attention has been posed to reproduce the discontinuities in the real blood vessels network, due to the non-uniform distribution of the contrast fluid (the blood stains by the contrast agent). Secondly, the availability of controllable artificial data is used to train and test a deep learning architecture, to verify its ability to perform semantic segmentation and reconstruct the discontinuities. The results will pro- vide relevant information for further research on real data and pose the bases for wider analysis and discussions. Di↵erent extension can be implemented to apply this method to other physiological structures and to other biomedical imaging techniques. The proposed method has the potentially to completely eliminates the dependence on real data for many biomedical image analysis.
Il sangue supporta lo scambio di ossigeno, sostanze nutritive e minerali tra i diversi tessuti del corpo umano. Una visualizzazione 3D dettagliata della completa rete vascolare fornisce informazioni utili sulla struttura, sull’organizzazione vascolare e sulla distribuzione del flusso sanguigno. Per isolare la rete 3D di vasi sanguigni dallo sfondo, è necessario eseguire una segmentazione semantica su immagini CE-CT, macchiate con un nuovo agente di contrasto (polyoxometalate). Gli approcci di deep learning per la segmentazione semantica sono promettenti e risolvono i problemi presenti nei metodi manuali e semi-automatici (tempo di computazione, mancanza di valutatori esperti, propensione agli errori, implementazione onerosa in un approccio 3D completo). Tuttavia, l’assenza di datasets pre-esistenti di immagini già segmentate, impedisce la fase di addestramento di queste reti neurali. La soluzione del progetto di tesi magistrale implementata per risolvere questi problemi è strutturata in due fasi differenti ma strettamente dipendenti. In primo luogo, è stato sviluppato un generatore 3D di immagini sintetiche, per riprodurre una rete realistica di vasi sanguigni 3D; lo scopo è di non essere più dipendenti dai dati reali. Le immagini sono state generate consentendo di creare un set di dati di grandi dimensioni con immagini artificiali e le loro coppie già segmentate, in modo da poterle impiegare durante l’addestramento delle reti neurali. Particolare attenzione è stata posta alle discontinuità presenti nella rete dei vasi sanguigni reali, dovute alla distribuzione non uniforme del liquido a contrasto. In secondo luogo, la disponibilità di dati artificiali è stata utilizzata per addestrare e valutare un’architettura di deep learning, per verificarne l’abilità di eseguire segmentazione semantica e ricostruire le discontinuità. I risultati forniscono informazioni pertinenti per ulteriori ricerche su dati reali e pongono le basi per applicare questo metodo ad altre strutture fisiologiche e ad altre tecniche di imaging biomedicale. Il metodo proposto ha la potenzialità di eliminare completamente la dipendenza dai dati reali.
Segmentation and branching analysis of the 3D blood vessel network in contrast-enhanced microCT images
POLI, SIMONE
2018/2019
Abstract
Blood supports the exchange of oxygen, nutrients and minerals be- tween di↵erent tissues of the human body. A detailed 3D visualization of the total vascular network provides useful information about the vascular structure and organization and the distribution of the blood flow. To isolate the 3D vasculature from the background, semantic segmentation must be performed on CE-CT images, stained with a novel contrast agent (polyoxometalate) that highlight the blood ves- sels. Deep learning approaches for semantic segmentation are promis- ing and solve the problems present in the manual and semi-automated methods (time-demanding, lack of experts, prone to errors, onerous to implement in a full 3D approach). However, the absence of an existing datasets of already segmented images, prevents the training phase. To solve these issues, the master thesis project is composed by two di↵er- ent but dependent steps: first, an artificial 3D synthetic data generator is developed, to reproduce realistic 3D blood vessels network, in order not to be dependent anymore from the real data. The images are gen- erated allowing to build large datasets with synthetic images and their already segmented pairs, to be used during neural network training. Particular attention has been posed to reproduce the discontinuities in the real blood vessels network, due to the non-uniform distribution of the contrast fluid (the blood stains by the contrast agent). Secondly, the availability of controllable artificial data is used to train and test a deep learning architecture, to verify its ability to perform semantic segmentation and reconstruct the discontinuities. The results will pro- vide relevant information for further research on real data and pose the bases for wider analysis and discussions. Di↵erent extension can be implemented to apply this method to other physiological structures and to other biomedical imaging techniques. The proposed method has the potentially to completely eliminates the dependence on real data for many biomedical image analysis.| File | Dimensione | Formato | |
|---|---|---|---|
|
Master_Thesis__SIMONE_POLI.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Testo della tesi
Dimensione
29.65 MB
Formato
Adobe PDF
|
29.65 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/10589/150969