The rapid adoption of artificial intelligence (AI) in oncological medical imaging has enabled the extraction of high-dimensional imaging biomarkers from routine clinical acquisitions, with machine learning (ML) and deep learning (DL) models often achieving expert-level performance in tasks such as tumor characterization, screening, and risk stratification. Despite these advances, clinical translation remains limited by the black-box nature of state-of-the-art models and by the absence of rigorous, quantitative, and standardized validation frameworks, particularly in data-limited and heterogeneous clinical settings. This PhD thesis proposes a unifying cross-domain validation paradigm for Explainable Artificial Intelligence (XAI), grounded in the systematic use of High Energy Physics (HEP) as a causally interpretable and quantitatively controlled testbed for the evaluation of explainability methods applied to medical imaging in oncology. The central hypothesis is that HEP, with its well-defined physical laws, large-scale simulated datasets, and explicit ground truth, provides an ideal surrogate environment to assess the reliability, robustness, and fidelity of AI and XAI techniques that cannot be objectively validated in clinical data alone. The research follows a multi-stage and multi-modality framework across radiomics, synthetic data generation, and transformer-based pipelines, relying on imaging datasets retrieved from The Cancer Imaging Archive (TCIA), with a primary focus on lung cancer. The proposed radiomics-based ML models integrated post-hoc explainability through SHAP to support feature selection and model interpretation for CT-based lung cancer characterization, while DL pipelines exploited pretrained Vision Transformers to investigate attention-based interpretability in low-dose CT screening. The explainability strategies were transferred to the HEP domain, where tabular and image-based datasets were constructed to mirror key characteristics of medical data, including sample size, class imbalance, and dimensionality. Within this controlled setting, XAI methods were quantitatively validated against known HEP ground truth, enabling large-scale reproducible assessments of explanation stability and faithfulness. Finally, the technically validated models were transferred back to the medical domain, where their outputs were assessed through complementary biological and clinical analyses, including correlations with RNA-sequencing data and expert radiological annotations. By integrating HEP technical validation with biological and clinical validation, this thesis establishes a principled framework for the development and assessment of explainable AI models in medical imaging, contributing toward more transparent, robust, and clinically meaningful AI-driven decision support systems in oncology.
La crescente diffusione dell’Intelligenza Artificiale (IA) nell’imaging medico oncologico ha permesso l'estrazione di biomarcatori d’immagine ad alta dimensionalità da acquisizioni cliniche di routine, con modelli di Machine Learning (ML) e Deep Learning (DL) che raggiungono frequentemente prestazioni paragonabili a quelle degli esperti in termini di caratterizzazione del tumore, screening e stratificazione del rischio. Nonostante questi progressi, l’adozione nella pratica clinica rimane limitata dalla natura black-box dei modelli allo stato dell'arte e dall'assenza di validation frameworks rigorosi, quantitativi e standardizzati, in particolare in contesti clinici caratterizzati da eterogeneità e scarsità di dati. Questa tesi di dottorato propone un paradigma unificato di validazione multi dominio per l'Intelligenza Artificiale Spiegabile (XAI), basato sull'uso sistematico della Fisica delle Alte Energie (HEP) come testbed causalmente interpretabile e quantitativamente controllato per la valutazione dei metodi di spiegabilità applicati all'imaging medico in ambito oncologico. L'ipotesi centrale è che la HEP, con le sue leggi fisiche ben definite, dataset simulati su larga scala e una ground truth esplicita, fornisca un ambiente surrogato ideale per valutare l’affidabilità,la stabilità e l’accuratezza delle tecniche di AI e XAI che non possono essere validate oggettivamente solo sui dati clinici. Lo studio segue un framework multi-dominio e modalità che integra pipeline di radiomica, generazione di dati sintetici e architetture basate su Transformer. Utilizzando dataset provenienti da The Cancer Imaging Archive (TCIA), con un focus primario sul tumore al polmone, sono stati sviluppati modelli di ML radiomici che integrano l’explainability post-hoc tramite SHAP per supportare la selezione delle feature e l’interpretazione del modello. Parallelamente, sono state implementate pipeline di DL, che sfruttano Vision Transformers pre-addestrati, al fine di investigare l’interpretabilità basata sui meccanismi di attenzione nel contesto dello screening con tomografia computerizzata (CT) a basso dosaggio (LDCT). Le strategie di spiegabilità sono state trasferite al dominio della HEP, dove sono stati costruiti dataset tabulari e basati su immagini per rispecchiare le caratteristiche chiave dei dati medici, tra cui la numerosità campionaria e lo sbilanciamento tra le classi. All'interno di questo contesto controllato, i metodi di XAI sono stati validati quantitativamente rispetto alla ground truth nota della HEP, consentendo valutazioni riproducibili su larga scala dell’affidabilità e della coerenza delle spiegazioni. Infine, i modelli tecnicamente validati sono stati ritrasferiti al dominio medico, dove i loro output sono stati valutati attraverso analisi biologiche e cliniche complementari, incluse correlazioni con dati di RNA-sequencing e annotazioni radiologiche di esperti. Integrando la validazione tecnica in ambito HEP con la validazione biologica e clinica, questa tesi stabilisce un framework metodologico per lo sviluppo e la valutazione di modelli di AI spiegabili nell'imaging medico, contribuendo allo sviluppo di sistemi di supporto alle decisioni guidati dall'AI più trasparenti, robusti e clinicamente significativi in oncologia.
Explainable artificial intelligence through high energy physics data for medical imaging in oncology
MONTELEONE, MARIAGRAZIA
2026
Abstract
The rapid adoption of artificial intelligence (AI) in oncological medical imaging has enabled the extraction of high-dimensional imaging biomarkers from routine clinical acquisitions, with machine learning (ML) and deep learning (DL) models often achieving expert-level performance in tasks such as tumor characterization, screening, and risk stratification. Despite these advances, clinical translation remains limited by the black-box nature of state-of-the-art models and by the absence of rigorous, quantitative, and standardized validation frameworks, particularly in data-limited and heterogeneous clinical settings. This PhD thesis proposes a unifying cross-domain validation paradigm for Explainable Artificial Intelligence (XAI), grounded in the systematic use of High Energy Physics (HEP) as a causally interpretable and quantitatively controlled testbed for the evaluation of explainability methods applied to medical imaging in oncology. The central hypothesis is that HEP, with its well-defined physical laws, large-scale simulated datasets, and explicit ground truth, provides an ideal surrogate environment to assess the reliability, robustness, and fidelity of AI and XAI techniques that cannot be objectively validated in clinical data alone. The research follows a multi-stage and multi-modality framework across radiomics, synthetic data generation, and transformer-based pipelines, relying on imaging datasets retrieved from The Cancer Imaging Archive (TCIA), with a primary focus on lung cancer. The proposed radiomics-based ML models integrated post-hoc explainability through SHAP to support feature selection and model interpretation for CT-based lung cancer characterization, while DL pipelines exploited pretrained Vision Transformers to investigate attention-based interpretability in low-dose CT screening. The explainability strategies were transferred to the HEP domain, where tabular and image-based datasets were constructed to mirror key characteristics of medical data, including sample size, class imbalance, and dimensionality. Within this controlled setting, XAI methods were quantitatively validated against known HEP ground truth, enabling large-scale reproducible assessments of explanation stability and faithfulness. Finally, the technically validated models were transferred back to the medical domain, where their outputs were assessed through complementary biological and clinical analyses, including correlations with RNA-sequencing data and expert radiological annotations. By integrating HEP technical validation with biological and clinical validation, this thesis establishes a principled framework for the development and assessment of explainable AI models in medical imaging, contributing toward more transparent, robust, and clinically meaningful AI-driven decision support systems in oncology.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD_Thesis_MariagraziaMonteleone_final.pdf
accesso solo da BNCF e BNCR
Licenza:
Tutti i diritti riservati
Dimensione
10.49 MB
Formato
Adobe PDF
|
10.49 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/372135
URN:NBN:IT:POLIMI-372135