Synchrotron Radiation micro-Computed Tomography (SR-microCT) offers a powerful view of bone microstructure, making it possible to observe and quantify small anatomical features that are otherwise difficult to study in three dimensions. Existing related work provides specialized SR-microCT segmentation and morphometry methods, bioimage-analysis workflow platforms, and emerging biomedical agents, but these directions have not yet converged into an integrated framework for traceable SR-microCT bone-analysis workflows. Reconstructed volumes must be appropriately processed to extract interpretable results, but this process often depends on heterogeneous tools, large files, manual parameter choices, and fragile passages between software environments. As a result, researchers interested in biological interpretation are often required to manage a demanding computational workflow before the data can be meaningfully analyzed. This thesis introduces OsteoGraph, a modular multi-agent framework for quantitative SR-microCT bone-image analysis. The system autonomously coordinates the passage from a researcher's analytical request to a sequence of tool-based operations. A Large Language Model (LLM) interprets the request, selects specialist components, invokes computational tools, and preserves the artifacts produced along the way. The resulting outputs remain connected to the operations that generated them. This makes it possible to inspect the analytical process as a whole, from the initial request and tool calls to the intermediate artifacts and final results, rather than treating the answer as a standalone conversational response. The contribution of OsteoGraph lies in making the analytical process more accessible while keeping its computational path visible. The framework is evaluated on representative SR-microCT analysis tasks spanning data handling, preprocessing, segmentation, quantitative analysis, and classification-oriented comparison. The evaluation shows that OsteoGraph can coordinate multi-step workflows over real image-derived artifacts, preserve intermediate outputs, and return structured results across the main stages of the analysis pipeline. The remaining limitations are concentrated in routing and tool sequencing, where errors in selecting or ordering specialist operations can affect the continuity of the workflow. OsteoGraph helps researchers move from scientific intent to documented computational evidence while preserving the role of tools and human judgement.
La microtomografia computerizzata a radiazione di sincrotrone (SR-microCT) offre una prospettiva dettagliata della microstruttura ossea, rendendo possibile osservare e quantificare caratteristiche anatomiche microscopiche che sarebbero altrimenti difficili da studiare in tre dimensioni. La letteratura propone metodi specializzati per la segmentazione e la morfometria in SR-microCT, piattaforme per analisi di bioimmagini e agenti biomedicali emergenti. Tuttavia, questi sforzi non sono ancora confluiti in un framework integrato per workflow tracciabili di analisi ossea basata su SR-microCT. I volumi ricostruiti devono essere opportunamente processati per estrarre risultati interpretabili, ma questo processo dipende spesso da strumenti eterogenei, file di grandi dimensioni, scelte manuali dei parametri e passaggi fragili tra ambienti software. Di conseguenza, i ricercatori che vogliono dare un'interpretazione biologica sono spesso costretti a gestire complessità computazionale impegnativa prima che i dati possano essere analizzati in modo significativo. Questa tesi introduce OsteoGraph, un framework multi-agente modulare per l'analisi quantitativa di immagini ossee SR-microCT. Il sistema coordina autonomamente il passaggio dalla richiesta analitica del ricercatore a una sequenza di operazioni basate su strumenti. Un Large Language Model (LLM) interpreta la richiesta, seleziona degli specialisti, invoca strumenti computazionali e preserva gli artefatti prodotti lungo il percorso. Gli output risultanti rimangono collegati alle operazioni che li hanno generati. Questo rende possibile ispezionare il processo analitico nel suo insieme, dalla richiesta iniziale e dalle chiamate agli strumenti fino agli artefatti intermedi e ai risultati finali, invece di trattare la risposta come un semplice output conversazionale autonomo. Il contributo di OsteoGraph consiste nel rendere il processo analitico più accessibile mantenendone visibile il percorso computazionale. Il framework viene valutato su task rappresentativi di analisi SR-microCT che includono gestione dei dati, preprocessing, segmentazione, analisi quantitativa e confronto orientato alla classificazione. La valutazione mostra che OsteoGraph può coordinare workflow multi-step su artefatti derivati da immagini reali, preservare gli output intermedi e restituire risultati strutturati attraverso le principali fasi della pipeline di analisi. Le limitazioni rimanenti si concentrano nel routing e nel sequenziamento degli strumenti, dove errori nella selezione o nell'ordinamento delle operazioni specialistiche possono influire sulla continuità del workflow. OsteoGraph aiuta i ricercatori a passare dall'intento scientifico a evidenze documentate, preservando al contempo il ruolo degli strumenti e del giudizio umano.
OsteoGraph: a multi-agent tool orchestration framework for quantitative bone microsrtucture analysis
Giannetto, Matteo
2025/2026
Abstract
Synchrotron Radiation micro-Computed Tomography (SR-microCT) offers a powerful view of bone microstructure, making it possible to observe and quantify small anatomical features that are otherwise difficult to study in three dimensions. Existing related work provides specialized SR-microCT segmentation and morphometry methods, bioimage-analysis workflow platforms, and emerging biomedical agents, but these directions have not yet converged into an integrated framework for traceable SR-microCT bone-analysis workflows. Reconstructed volumes must be appropriately processed to extract interpretable results, but this process often depends on heterogeneous tools, large files, manual parameter choices, and fragile passages between software environments. As a result, researchers interested in biological interpretation are often required to manage a demanding computational workflow before the data can be meaningfully analyzed. This thesis introduces OsteoGraph, a modular multi-agent framework for quantitative SR-microCT bone-image analysis. The system autonomously coordinates the passage from a researcher's analytical request to a sequence of tool-based operations. A Large Language Model (LLM) interprets the request, selects specialist components, invokes computational tools, and preserves the artifacts produced along the way. The resulting outputs remain connected to the operations that generated them. This makes it possible to inspect the analytical process as a whole, from the initial request and tool calls to the intermediate artifacts and final results, rather than treating the answer as a standalone conversational response. The contribution of OsteoGraph lies in making the analytical process more accessible while keeping its computational path visible. The framework is evaluated on representative SR-microCT analysis tasks spanning data handling, preprocessing, segmentation, quantitative analysis, and classification-oriented comparison. The evaluation shows that OsteoGraph can coordinate multi-step workflows over real image-derived artifacts, preserve intermediate outputs, and return structured results across the main stages of the analysis pipeline. The remaining limitations are concentrated in routing and tool sequencing, where errors in selecting or ordering specialist operations can affect the continuity of the workflow. OsteoGraph helps researchers move from scientific intent to documented computational evidence while preserving the role of tools and human judgement.| File | Dimensione | Formato | |
|---|---|---|---|
|
full tesi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: tesi di laurea
Dimensione
3.11 MB
Formato
Adobe PDF
|
3.11 MB | Adobe PDF | Visualizza/Apri |
|
executive summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: executive summary
Dimensione
234.8 kB
Formato
Adobe PDF
|
234.8 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/10589/260300