Radiotherapy is a cornerstone in cancer treatment, used in over half of all cancer patients. However, full benefit is often compromised by unintended radiation exposure to nearby healthy organs, leading to acute and chronic side effects that significantly impact patients' quality of life. To assess and help prevent treatment-related toxicities, predictive models—known as Normal Tissue Complication Probability (NTCP) models—are employed to link radiation dose with observed toxicities. Among these, models based on the Equivalent Uniform Dose (EUD) have been widely adopted, but suffer from a key limitation: they ignore the spatial distribution of dose within organs, which may play a critical role in the development of toxicities. To address this limitation, this thesis introduces a novel framework that leverages Functional Principal Component Analysis (FPCA) to capture spatial variation in dose distributions. FPCA extracts the dominant patterns of dose across patients and represents each individual distribution as a weighted combination of these patterns. The resulting patient-specific scores are then used as covariates in Cox Proportional Hazards (CPH) models to estimate the risk of developing specific side effects over time. A central mathematical result supports the method’s validity and interpretability: for each patient, the risk score produced by the CPH model is mathematically equivalent to the L^2 inner product between their centered dose distribution and a toxicity-specific spatial risk pattern derived from the model itself. This formulation allows for direct visualization of high-risk dose regions, offering clinicians interpretable, anatomy-specific insights into the spatial determinants of radiation-induced side effects. The method was first developed using rectal dose distributions from the REQUITE prostate dataset and validated on the Micro-Learner prostate cohort, confirming its robustness. It was then applied to the Micro-Learner head-and-neck study, demonstrating adaptability across clinical sites. Across all applications, FPCA-based models revealed anatomically distinct high-risk dose regions predictive of toxicity and outperformed traditional EUD-based approaches. Beyond improved predictive performance, the FPCA approach enhances our understanding of the spatial mechanisms underlying radiation-induced toxicities, offering valuable insights into the anatomical and dosimetric factors that drive treatment-related side effects.
La radioterapia rappresenta una componente fondamentale nel trattamento del cancro, impiegata in oltre la metà dei pazienti oncologici. Tuttavia, il pieno beneficio della terapia è spesso limitato dall’esposizione indesiderata di tessuti sani adiacenti, che può causare effetti collaterali sia acuti che cronici, compromettendo significativamente la qualità di vita dei pazienti. Per prevedere e ridurre le tossicità correlate al trattamento, si utilizzano modelli predittivi noti come Modelli di Probabilità di Complicanze nei Tessuti Normali (Normal Tissue Complication Probability Models, NTCPM), che mettono in relazione la dose di radiazione somministrata con la comparsa degli effetti avversi. Tra questi, i modelli basati sulla Dose Uniforme Equivalente (Equivalent Uniform Dose, EUD) sono largamente impiegati, ma presentano un’importante limitazione: non considerano la distribuzione spaziale della dose all’interno degli organi, elemento che potrebbe influenzare in modo decisivo lo sviluppo delle tossicità. Per rispondere a questa criticità, questa tesi propone un nuovo approccio che utilizza l’Analisi delle Componenti Principali Funzionali (Functional Principal Component Analysis, FPCA) per cogliere la variazione spaziale nelle distribuzioni di dose. FPCA individua i pattern dominanti di distribuzione della dose tra i pazienti e rappresenta ogni singola distribuzione come una combinazione pesata di tali pattern. I punteggi specifici per paziente ottenuti da questa analisi vengono poi utilizzati come covariate in modelli di Rischio Proporzionale di Cox (Cox Proportional Hazards, CPH) per stimare il rischio di sviluppare effetti collaterali nel tempo. Un risultato matematico centrale garantisce la validità e l’interpretabilità del metodo: per ogni paziente, il punteggio di rischio calcolato dal modello CPH è matematicamente equivalente al prodotto interno L^2 tra la distribuzione di dose centrata del paziente e un pattern spaziale di rischio specifico della tossicità, derivato direttamente dal modello. Questa formulazione permette di visualizzare le aree a maggior rischio di dose, offrendo ai clinici una comprensione interpretabile e anatomica dei fattori spaziali che determinano gli effetti collaterali indotti dalla radioterapia. Il metodo è stato inizialmente sviluppato utilizzando le distribuzioni di dose rettali del dataset REQUITE per il tumore alla prostata e validato sul coorte Micro-Learner prostata, confermandone la robustezza. Successivamente è stato applicato allo studio Micro-Learner per i tumori testa-collo, dimostrando la sua adattabilità a diversi contesti clinici. In tutte le applicazioni, i modelli basati su FPCA hanno evidenziato regioni anatomiche ad alto rischio di tossicità e hanno superato le prestazioni dei modelli tradizionali basati su EUD. Oltre a migliorare le capacità predittive, l’approccio FPCA contribuisce ad approfondire la comprensione dei meccanismi spaziali alla base delle tossicità indotte dalla radioterapia, offrendo preziose informazioni sui fattori anatomici e dosimetrici che guidano gli effetti collaterali correlati al trattamento.
Functional principal component analysis of 2D radiotherapy dose maps to identify toxicity risk patterns
VILLA, GIANLUCA
2024/2025
Abstract
Radiotherapy is a cornerstone in cancer treatment, used in over half of all cancer patients. However, full benefit is often compromised by unintended radiation exposure to nearby healthy organs, leading to acute and chronic side effects that significantly impact patients' quality of life. To assess and help prevent treatment-related toxicities, predictive models—known as Normal Tissue Complication Probability (NTCP) models—are employed to link radiation dose with observed toxicities. Among these, models based on the Equivalent Uniform Dose (EUD) have been widely adopted, but suffer from a key limitation: they ignore the spatial distribution of dose within organs, which may play a critical role in the development of toxicities. To address this limitation, this thesis introduces a novel framework that leverages Functional Principal Component Analysis (FPCA) to capture spatial variation in dose distributions. FPCA extracts the dominant patterns of dose across patients and represents each individual distribution as a weighted combination of these patterns. The resulting patient-specific scores are then used as covariates in Cox Proportional Hazards (CPH) models to estimate the risk of developing specific side effects over time. A central mathematical result supports the method’s validity and interpretability: for each patient, the risk score produced by the CPH model is mathematically equivalent to the L^2 inner product between their centered dose distribution and a toxicity-specific spatial risk pattern derived from the model itself. This formulation allows for direct visualization of high-risk dose regions, offering clinicians interpretable, anatomy-specific insights into the spatial determinants of radiation-induced side effects. The method was first developed using rectal dose distributions from the REQUITE prostate dataset and validated on the Micro-Learner prostate cohort, confirming its robustness. It was then applied to the Micro-Learner head-and-neck study, demonstrating adaptability across clinical sites. Across all applications, FPCA-based models revealed anatomically distinct high-risk dose regions predictive of toxicity and outperformed traditional EUD-based approaches. Beyond improved predictive performance, the FPCA approach enhances our understanding of the spatial mechanisms underlying radiation-induced toxicities, offering valuable insights into the anatomical and dosimetric factors that drive treatment-related side effects.| File | Dimensione | Formato | |
|---|---|---|---|
|
Functional Principal Component Analysis of 2D Radiotherapy Dose Maps to Identify Toxicity Risk Patterns - Gianluca Villa Master Thesis.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Functional Principal Component Analysis of 2D Radiotherapy Dose Maps to Identify Toxicity Risk Patterns - Gianluca Villa Master Thesis
Dimensione
8.46 MB
Formato
Adobe PDF
|
8.46 MB | Adobe PDF | Visualizza/Apri |
|
Functional Principal Component Analysis of 2D Radiotherapy Dose Maps to Identify Toxicity Risk Patterns - Gianluca Villa Executive Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Functional Principal Component Analysis of 2D Radiotherapy Dose Maps to Identify Toxicity Risk Patterns - Gianluca Villa Executive Summary
Dimensione
825.8 kB
Formato
Adobe PDF
|
825.8 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/10589/240747