The assessment of the integrity of genetic material is one of the fundamental pillars of modern toxicology, biomedical diagnostics and cancer research. Indeed, chemical and/or physical agents can cause damage to genetic material through endogenous or exogenous causes. Among the biochemical methods used to quantitatively estimate single- or double-strand breaks in DNA, the comet assay occupies a prominent position thanks to its high sensitivity, versatility and relatively low operating cost. This technique is based on the negative electrical charge of DNA, which, when subjected to an electric field in an agarose gel, migrates towards the anode; intact nuclei will migrate in a compact manner, whilst nuclei with DNA subject to strand breaks will migrate in a way that produces the characteristic morphological structure resembling the shape of a comet, hence the name of the technique. Using fluorochromes, it is then possible to label the DNA and detect quantities of material in the ‘head’ and in the ‘tail’, which may be more or less extended. Despite its effectiveness being widely documented in the literature, the conventional analytical methodology suffers from significant operational limitations. Traditionally, the classification of the level of genotoxic damage is carried out by visual inspection under a fluorescence microscope by a human operator, who assigns each individual cell to one of the five damage classes defined by international protocols. This manual approach introduces a high degree of subjectivity, particularly when identifying the morphological boundaries between adjacent mild damage classes. Furthermore, the process is very time-consuming and is hampered by the operator’s visual fatigue, limiting the statistical reproducibility of the tests and acting as a clear bottleneck for high-throughput laboratory activities. These techniques are now complemented by quantitative image analysis methods, which nevertheless suffer from a high level of inter-operator subjectivity. It is within this technological and clinical context that this thesis is situated, developed as part of computational and experimental internship activities. The primary objective of this thesis is the design, optimisation and validation of an automated software pipeline written in Python, based on the use of Artificial Intelligence algorithms and, specifically, Convolutional Neural Networks, for the objective and automatic classification of images of nuclei undergoing the comet assay. The workflow involved a combination of laboratory-based experimental work and computational modelling. Part of the project was dedicated to conducting direct optical acquisition sessions under the microscope to expand the dataset, enabling the model to be trained on a solid statistical basis of over a thousand samples. The results obtained were finally analysed not only using standard quantitative metrics, but also through a qualitative approach, aimed at investigating the biological and optical causes underlying the model’s misclassifications.
La valutazione dell’integrità del materiale genetico rappresenta uno dei pilastri fondamentali della tossicologia moderna, della diagnostica biomedica e della ricerca oncologica. Difatti, agenti chimici e/o fisici, possono indurre il danneggiamento del materiale genetico per cause endogene o esogene. Tra le metodologie biochimiche impiegate per stimare quantitativamente le rotture a singolo o doppio filamento del DNA, il saggio della cometa (o Comet assay), occupa una posizione di preminenza grazie alla sua elevata sensibilità, versatilità e relativo basso costo operativo. Questa tecnica si basa sulla carica elettrica negativa del DNA, che, sottoposto ad un campo elettrico in un gel di agarosio, migra verso l’anodo; nuclei integri migreranno in ma- niera compatta, mentre nuclei con DNA soggetto a rotture del filamento migreranno determinando la caratteristica struttura morfologica che ricorda la forma di una cometa, da cui il nome della tecnica. Mediante fluorocromi è poi possibile marcare il DNA e rilevare quantitativi di materiale nella “testa” e nella “coda”, più o meno estesa. Nonostante l’efficacia ampiamente documentata in letteratura, la metodologia di analisi convenzionale risente di significativi limiti operativi. Tradizionalmente, la classificazione del livello di danno genotossico viene delegata all’ispezione visiva al microscopio a fluorescenza da parte di un operatore umano, il quale assegna ogni singola cellula a una delle cinque classi di danno previste dai protocolli internazionali. Questo approccio manuale introduce una forte componente di soggettività, special- mente nella rilevazione dei confini morfologici tra classi adiacenti lievi. Inoltre, il processo risulta molto lungo e penalizzato dall’affaticamento visivo dell’operatore, limitando la riproducibilità statistica dei test e configurandosi come un evidente collo di bottiglia per le attività di laboratorio ad alta produttività. Queste tecniche sono oggi affiancate da metodi quantiativi di analisi di immagine, che comunque risentono di un elevato livello di soggettività interoperatore. In questo scenario tecnologico e clinico si inserisce il presente lavoro di tesi, sviluppato nell’ambito delle attività di tirocinio computazionale e sperimentale. L’obiettivo primario dell’elaborato consiste nella progettazione, ottimizzazione e validazione di una pipeline software automatizzata in linguaggio Python, basata sull’impiego di algoritmi di Intelligenza Artificiale e, nello specifico, di Reti Neurali Convoluzionali, per la classificazione oggettiva e automatica delle immagini di nuclei sottoposti al test della cometa. Il flusso di lavoro ha previsto una combinazione tra la fase sperimentale di labora- torio e la modellizzazione informatica. Una parte del progetto è stata dedicata alla conduzione di sessioni di acquisizione ottica diretta al microscopio per l’espansione del dataset, permettendo di addestrare l’architettura su una solida base statistica di oltre mille campioni. I risultati ottenuti sono stati infine analizzati non solo mediante metriche quantitative standard, ma anche attraverso un approccio qualitativo, volto a investigare le cause biologiche e ottiche alla base delle misclassificazioni del modello.
Sviluppo di algoritmi ML per la classificazione di immagini ottenute da test per la valutazione del danno genotossico
MORIGI, MARIA CECILIA
2025/2026
Abstract
The assessment of the integrity of genetic material is one of the fundamental pillars of modern toxicology, biomedical diagnostics and cancer research. Indeed, chemical and/or physical agents can cause damage to genetic material through endogenous or exogenous causes. Among the biochemical methods used to quantitatively estimate single- or double-strand breaks in DNA, the comet assay occupies a prominent position thanks to its high sensitivity, versatility and relatively low operating cost. This technique is based on the negative electrical charge of DNA, which, when subjected to an electric field in an agarose gel, migrates towards the anode; intact nuclei will migrate in a compact manner, whilst nuclei with DNA subject to strand breaks will migrate in a way that produces the characteristic morphological structure resembling the shape of a comet, hence the name of the technique. Using fluorochromes, it is then possible to label the DNA and detect quantities of material in the ‘head’ and in the ‘tail’, which may be more or less extended. Despite its effectiveness being widely documented in the literature, the conventional analytical methodology suffers from significant operational limitations. Traditionally, the classification of the level of genotoxic damage is carried out by visual inspection under a fluorescence microscope by a human operator, who assigns each individual cell to one of the five damage classes defined by international protocols. This manual approach introduces a high degree of subjectivity, particularly when identifying the morphological boundaries between adjacent mild damage classes. Furthermore, the process is very time-consuming and is hampered by the operator’s visual fatigue, limiting the statistical reproducibility of the tests and acting as a clear bottleneck for high-throughput laboratory activities. These techniques are now complemented by quantitative image analysis methods, which nevertheless suffer from a high level of inter-operator subjectivity. It is within this technological and clinical context that this thesis is situated, developed as part of computational and experimental internship activities. The primary objective of this thesis is the design, optimisation and validation of an automated software pipeline written in Python, based on the use of Artificial Intelligence algorithms and, specifically, Convolutional Neural Networks, for the objective and automatic classification of images of nuclei undergoing the comet assay. The workflow involved a combination of laboratory-based experimental work and computational modelling. Part of the project was dedicated to conducting direct optical acquisition sessions under the microscope to expand the dataset, enabling the model to be trained on a solid statistical basis of over a thousand samples. The results obtained were finally analysed not only using standard quantitative metrics, but also through a qualitative approach, aimed at investigating the biological and optical causes underlying the model’s misclassifications.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Maria_Cecilia_Morigi.pdf
accesso aperto
Descrizione: caricamento Tesi Morigi Maria Cecilia ingegneria biomedica s1109081
Dimensione
4.72 MB
Formato
Adobe PDF
|
4.72 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/27257