Prostate cancer is a leading cause of male mortality worldwide. Early diagnosis is essential to ensure adequate treatment and improve therapeutic outcomes. Multiparametric magnetic resonance imaging (mpMRI) is one of the most promising techniques for the identification and characterization of prostatic lesions, thanks to its ability to provide detailed images of the prostate and suspicious areas of malignancy. A scoring system such as PI-RADS (Prostate Imaging Reporting and Data System) is used to classify the severity of lesions, but the interpretation of these images is complex and subjective, making an automated approach necessary to support clinicians in the diagnosis. In this thesis work, carried out in collaboration with the IRCCS CROB of Rionero in Vulture (PZ) in the field of radiomic analysis, machine learning models for the prediction of the PI-RADS score have been developed and implemented, using radiomic features extracted from masks of the ROIs of the lesions obtained from mpMRI images of the prostate, first acquired with standard scanning parameters (TE and TR) and then modified. The implemented classification models include logistic regression, Random Forest, Support Vector Machine (SVM) and a deep neural network, with the aim of not only evaluating their ability to predict the PI-RADS score, but also comparing their performance to determine whether they can predict with greater accuracy on the modified dataset, compared to the standard one. The workflow includes data preprocessing, which includes the normalization of the features extracted from the images, followed by model training. The performances of the models were compared using evaluation metrics such as accuracy, precision, recall and F1-score. The obtained results show that the machine learning models are able to predict PI-RADS scores with good accuracy, with an improvement in performance when applied to the dataset of features extracted from the masks of the ROIs of the lesions obtained from the mpMRI images of the prostate acquired with standard parameters. However, the performances remain quite satisfactory even when the models are applied to the modified dataset.
Il carcinoma prostatico è una delle principali cause di mortalità maschile a livello mondiale. Una diagnosi precoce è fondamentale per garantire il trattamento adeguato e migliorare gli esiti terapeutici. La risonanza magnetica multiparametrica (mpMRI) è una delle tecniche più promettenti per l’identificazione e la caratterizzazione delle lesioni prostatiche, grazie alla sua capacità di fornire immagini dettagliate della prostata e delle aree sospette di malignità. Un sistema di scoring come il PI-RADS (Prostate Imaging Reporting and Data System) è utilizzato per classificare la gravità delle lesioni, ma l’interpretazione di queste immagini è complessa e soggettiva, rendendo necessario un approccio automatico per supportare i clinici nella diagnosi. Nel presente lavoro di tesi, realizzato in collaborazione con l’IRCCS CROB di Rionero in Vulture (PZ) nell’ambito dell’analisi radiomica, sono stati sviluppati e implementati modelli di machine learning per la predizione del punteggio PI-RADS, utilizzando caratteristiche radiomiche estratte da maschere delle ROI delle lesioni ottenute dalle immagini mpMRI della prostata, acquisite prima con parametri di scansione (TE e TR) standard e poi modificati. I modelli di classificazione implementati comprendono la regressione logistica, Random Forest, Support Vector Machine (SVM) e una rete neurale profonda, con l’obiettivo di non solo valutare la loro capacità di predire il punteggio PI-RADS, ma anche di confrontare le loro prestazioni per determinare se riescono a predire con maggiore accuratezza sul dataset modificato, rispetto a quello standard. Il flusso di lavoro comprende il preprocessing dei dati, che include la normalizzazione delle caratteristiche estratte dalle immagini, seguita poi dall’addestramento dei modelli. Le performance dei modelli sono state confrontate utilizzando metriche di valutazione come accuracy, precision, recall e F1-score. I risultati ottenuti mostrano che i modelli di machine learning sono in grado di predire con buona accuratezza i punteggi PI-RADS, con un miglioramento delle prestazioni quando sono applicati al dataset delle caratteristiche estratte dalle maschere delle ROI delle lesioni ottenute dalle immagini mpMRI della prostata acquisite con parametri standard. Tuttavia, le prestazioni restano comunque discretamente soddisfacenti anche quando i modelli sono applicati al dataset modificato.
MODELLAZIONE E SVILUPPO DI UN SISTEMA INTELLIGENTE PER LA CARATTERIZZAZIONE DEL TUMORE DELLA PROSTATA DA RISONANZA MAGNETICA
PITOIA, ROSALBA
2023/2024
Abstract
Prostate cancer is a leading cause of male mortality worldwide. Early diagnosis is essential to ensure adequate treatment and improve therapeutic outcomes. Multiparametric magnetic resonance imaging (mpMRI) is one of the most promising techniques for the identification and characterization of prostatic lesions, thanks to its ability to provide detailed images of the prostate and suspicious areas of malignancy. A scoring system such as PI-RADS (Prostate Imaging Reporting and Data System) is used to classify the severity of lesions, but the interpretation of these images is complex and subjective, making an automated approach necessary to support clinicians in the diagnosis. In this thesis work, carried out in collaboration with the IRCCS CROB of Rionero in Vulture (PZ) in the field of radiomic analysis, machine learning models for the prediction of the PI-RADS score have been developed and implemented, using radiomic features extracted from masks of the ROIs of the lesions obtained from mpMRI images of the prostate, first acquired with standard scanning parameters (TE and TR) and then modified. The implemented classification models include logistic regression, Random Forest, Support Vector Machine (SVM) and a deep neural network, with the aim of not only evaluating their ability to predict the PI-RADS score, but also comparing their performance to determine whether they can predict with greater accuracy on the modified dataset, compared to the standard one. The workflow includes data preprocessing, which includes the normalization of the features extracted from the images, followed by model training. The performances of the models were compared using evaluation metrics such as accuracy, precision, recall and F1-score. The obtained results show that the machine learning models are able to predict PI-RADS scores with good accuracy, with an improvement in performance when applied to the dataset of features extracted from the masks of the ROIs of the lesions obtained from the mpMRI images of the prostate acquired with standard parameters. However, the performances remain quite satisfactory even when the models are applied to the modified dataset.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Completa_Pitoia.pdf
File consultabile da postazioni biblioteca
Descrizione: TESI COMPLETA
Dimensione
4.34 MB
Formato
Adobe PDF
|
4.34 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/86631