Two-photon polymerization is an advanced bioprinting technique capable of fabricating high-resolution 3D microstructures, including constructs embedded with living cells. Despite its promise in tissue engineering, two-photon polymerization still lacks structured monitoring strategies capable of ensuring quality and reproducibility, particularly when cell-laden resins are involved. Imaging noise, morphological variability, and annotation complexity present substantial challenges to the development of data-driven control systems. This thesis presents a novel image-based framework for the continuous assessment of printability along the z-axis in multilayer two-photon polymerized constructs. An experimental campaign was conducted using a full-factorial Design of Experiments, systematically varying scan speed and laser power across 16 parameter combinations, each replicated three times. Experiments were performed for two material conditions: GelMA-based resin without cells and GelMA-based resin containing embedded C2C12 cells. For each print, a stack of 101 grayscale images - one per printed layer - was acquired for analysis. Upon visual inspection, the dataset encompassed a range of structural outcomes -including absence of printing, successful polymerization, and overexposure-related defects such as bubble formation. To systematically capture this variability, a quality scoring system ranging from 0 to 3 was defined. To develop a predictive model for model structural quality while minimizing annotation requirements, both supervised and semi-supervised learning strategies were implemented. The baseline model (M1) was a fully supervised ResNet-50, trained on the entire labeled dataset. The alternative (M2) was a two-stage pipeline consisting of a self-supervised inpainting pretraining followed by lightweight fine-tuning on manually labeled images. A third model (M3) tested transfer learning performance under constrained supervision. The results demonstrate that M2 achieves comparable accuracy to M1 while requiring significantly fewer labeled samples, and exhibits better generalization than M3. The integration of a Bayesian output head further enabled prediction uncertainty estimation, allowing quality assessment to be coupled with risk indicators. Results showed that the proposed pipeline enables layer-wise monitoring of structural quality during the printing process. In addition, by generating interpolated printability maps across the experimental space, it reveals trends in process stability over different parameter combinations and material conditions across layers. These maps facilitate the identification of optimal printability windows for various materials and layer depths, providing practitioners with valuable insights to select appropriate parameters tailored to specific conditions and to ensure consistent print quality throughout the build. Overall, this work represents the first 2PP monitoring framework suitable for biologically loaded constructs, combining interpretability, scalability, and real-time feasibility. It offers a foundation for adaptive control solutions in high-resolution biofabrication.
La polimerizzazione a due fotoni è una tecnica di biostampa avanzata in grado di realizzare microstrutture 3D ad alta risoluzione, comprese quelle contenenti cellule viventi. Nonostante il suo potenziale nel campo dell’ingegneria tissutale, la stampa a due fotoni è ancora priva di strategie di monitoraggio strutturate in grado di garantire qualità e riproducibilità, soprattutto nel caso di resine caricate con cellule. Il rumore di imaging, la variabilità morfologica e la complessità dell’annotazione rappresentano infatti sfide significative nello sviluppo di sistemi di controllo basati sui dati. Questa tesi presenta un framework innovativo basato su immagini per la valutazione continua della stampabilità lungo l’asse Z di costrutti ottenuti tramite 2PP multistrato. L’attività sperimentale è stata progettata tramite un Design of Experiments full-factorial, che ha previsto la variazione sistematica della velocità di scansione e della potenza laser su 16 combinazioni di parametri, ciascuna ripetuta tre volte. Gli esperimenti sono stati condotti per due condizioni di materiale: una resina a base di GelMA senza cellule e la stessa resina caricata con cellule C2C12. Per ogni stampa è stata acquisita una pila di 101 immagini in scala di grigi (una per ogni layer), poi utilizzate per l’analisi. L’ispezione visiva del dataset ha mostrato un’ampia varietà di risultati strutturali, tra cui assenza di stampa, polimerizzazione corretta e difetti da sovraesposizione come la formazione di bolle. Per descrivere sistematicamente questa variabilità, è stato definito un sistema di classificazione qualitativa da 0 a 3. Al fine di predire automaticamente la qualità strutturale riducendo il bisogno di etichettatura manuale, sono state implementate strategie di apprendimento supervisionato e semi-supervisionato. Il modello di riferimento (M1) è una ResNet-50 supervisionata, addestrata sull’intero dataset annotato. Il modello alternativo (M2) utilizza invece un pretraining auto-supervisionato tramite inpainting, seguito da un fine-tuning leggero su un set ridotto di immagini etichettate. Un terzo modello (M3) ha valutato le performance del transfer learning in condizioni di supervisione limitata. I risultati mostrano che M2 raggiunge un’accuratezza comparabile a M1 pur richiedendo molte meno etichette manuali, e generalizza meglio di M3. L’integrazione di una testa bayesiana in uscita ha inoltre permesso di stimare l’incertezza della predizione, associando alla valutazione qualitativa un indicatore di affidabilità. Il framework proposto consente un monitoraggio layer-by-layer della qualità strutturale durante la stampa. Inoltre, la generazione di mappe interpolanti della stampabilità nello spazio dei parametri sperimentali ha evidenziato tendenze nella stabilità del processo in funzione delle diverse combinazioni sperimentali e condizioni materiali. Tali mappe facilitano l’identificazione delle finestre ottimali di stampa, offrendo uno strumento utile alla scelta mirata dei parametri in funzione del materiale e della profondità, per garantire una qualità strutturale costante lungo tutta la stampa. Nel complesso, questo lavoro rappresenta il primo framework di monitoraggio per la stampa a due fotoni applicabile anche a materiali biologicamente attivi, combinando interpretabilità, scalabilità e applicabilità in tempo reale. Esso fornisce le basi per future soluzioni di controllo adattativo nella biofabbricazione ad alta risoluzione.
Predictive modeling and in-situ optimization of two-photon bioprinting using deep learning techniques
Bastianini, Beatrice
2024/2025
Abstract
Two-photon polymerization is an advanced bioprinting technique capable of fabricating high-resolution 3D microstructures, including constructs embedded with living cells. Despite its promise in tissue engineering, two-photon polymerization still lacks structured monitoring strategies capable of ensuring quality and reproducibility, particularly when cell-laden resins are involved. Imaging noise, morphological variability, and annotation complexity present substantial challenges to the development of data-driven control systems. This thesis presents a novel image-based framework for the continuous assessment of printability along the z-axis in multilayer two-photon polymerized constructs. An experimental campaign was conducted using a full-factorial Design of Experiments, systematically varying scan speed and laser power across 16 parameter combinations, each replicated three times. Experiments were performed for two material conditions: GelMA-based resin without cells and GelMA-based resin containing embedded C2C12 cells. For each print, a stack of 101 grayscale images - one per printed layer - was acquired for analysis. Upon visual inspection, the dataset encompassed a range of structural outcomes -including absence of printing, successful polymerization, and overexposure-related defects such as bubble formation. To systematically capture this variability, a quality scoring system ranging from 0 to 3 was defined. To develop a predictive model for model structural quality while minimizing annotation requirements, both supervised and semi-supervised learning strategies were implemented. The baseline model (M1) was a fully supervised ResNet-50, trained on the entire labeled dataset. The alternative (M2) was a two-stage pipeline consisting of a self-supervised inpainting pretraining followed by lightweight fine-tuning on manually labeled images. A third model (M3) tested transfer learning performance under constrained supervision. The results demonstrate that M2 achieves comparable accuracy to M1 while requiring significantly fewer labeled samples, and exhibits better generalization than M3. The integration of a Bayesian output head further enabled prediction uncertainty estimation, allowing quality assessment to be coupled with risk indicators. Results showed that the proposed pipeline enables layer-wise monitoring of structural quality during the printing process. In addition, by generating interpolated printability maps across the experimental space, it reveals trends in process stability over different parameter combinations and material conditions across layers. These maps facilitate the identification of optimal printability windows for various materials and layer depths, providing practitioners with valuable insights to select appropriate parameters tailored to specific conditions and to ensure consistent print quality throughout the build. Overall, this work represents the first 2PP monitoring framework suitable for biologically loaded constructs, combining interpretability, scalability, and real-time feasibility. It offers a foundation for adaptive control solutions in high-resolution biofabrication.File | Dimensione | Formato | |
---|---|---|---|
2025_07_Bastianini_Tesi_01.pdf
accessibile in internet per tutti
Descrizione: This work presents an image-based monitoring framework for two-photon polymerization processes, aimed at evaluating printability across layers. Using and comparing both supervised and semi-supervised models, including uncertainty-aware predictions, the system enables accurate classification of structural outcomes and supports quality assessment in cell-laden and cell-free materials. The proposed approach enhances process interpretability, minimizes manual labeling, and offers valuable insights for real-time control in high-resolution bioprinting.
Dimensione
7.37 MB
Formato
Adobe PDF
|
7.37 MB | Adobe PDF | Visualizza/Apri |
2025_07_Bastianini_Executive Summary_02.pdf
accessibile in internet per tutti
Descrizione: This work presents an image-based monitoring framework for two-photon polymerization processes, aimed at evaluating printability across layers. Using and comparing both supervised and semi-supervised models, including uncertainty-aware predictions, the system enables accurate classification of structural outcomes and supports quality assessment in cell-laden and cell-free materials. The proposed approach enhances process interpretability, minimizes manual labeling, and offers valuable insights for real-time control in high-resolution bioprinting.
Dimensione
862.71 kB
Formato
Adobe PDF
|
862.71 kB | 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/10589/240493