Thesis etd-09152025-111024 |
Link copiato negli appunti
Thesis type
Tesi di laurea magistrale
Author
BRADASCIO, LINDA DOMINIQUE
URN
etd-09152025-111024
Thesis title
Optimizing the clinical workflow of whole-body diffusion-weighted MRI in oncologic radiology with AI
Department
INGEGNERIA DELL'INFORMAZIONE
Course of study
INGEGNERIA BIOMEDICA
Supervisors
relatore Prof. Positano, Vincenzo
relatore Prof. Maes, Frederik
correlatore Dott. Vandecaveye, Vincent
relatore Prof. Maes, Frederik
correlatore Dott. Vandecaveye, Vincent
Keywords
- automatic segmentation
- classification
- classificazione
- CNN
- DW-MRI
- linfoma
- linfonodi
- lymph nodes
- lymphoma
- Random Forest
- segmentazione automatica
Graduation session start date
10/10/2025
Availability
Withheld
Release date
10/10/2095
Summary
La segmentazione e la classificazione dei linfonodi nelle immagini DW-MRI sono cruciali per la diagnosi e il monitoraggio del linfoma, tumore in crescita a livello globale. La DW-MRI, tecnica non invasiva, sfrutta la diversa diffusione delle molecole d’acqua per distinguere tessuti maligni, più densi e compatti, da quelli benigni, caratterizzati da maggiore libertà di movimento delle molecole. La segmentazione automatica, realizzata tramite reti neurali convoluzionali, riduce tempi ed errori rispetto ai metodi manuali. In questo studio, su 45 pazienti, è stato sviluppato un modello di segmentazione basato sull’architettura No New Net, con input combinati (ADC e b1000), ottenendo un Dice medio di 0.52, limitato da numero ridotto di pazienti, risoluzione non ottimale e immagini di ground truth parziali e imprecise. Per la classificazione, su 18 pazienti (1562 linfonodi benigni e 33 maligni) è stato impiegato un Random Forest addestrato su caratteristiche morfologiche e statistiche selezionate tramite PCA. Le performance hanno mostrato sensibilità 0.80, specificità 0.67 e accuratezza 0.74, ma lo squilibrio tra classi resta critico. Lo studio evidenzia il potenziale della DW-MRI in oncologia, sottolineando però la necessità di dataset più ampi e bilanciati e immagini di ground truth più accurate.
The segmentation and classification of lymph nodes in DW-MRI images are crucial for the diagnosis and monitoring of lymphoma, a tumor with a growing global incidence. DW-MRI, a non-invasive technique, exploits the different diffusion of water molecules to distinguish malignant tissues, which are denser and more compact, from benign ones, characterized by greater molecular mobility. Automatic segmentation, performed using convolutional neural networks, reduces time and errors compared to manual methods. In this study, on 45 patients, a segmentation model based on the No New Net architecture was developed, with combined inputs (ADC and b1000), achieving a mean Dice score of 0.52, limited by the small number of patients, suboptimal resolution and partially inaccurate ground truth images. For the classification, a Random Forest was trained on morphological and statistical features selected through PCA on 18 patients (1562 benign and 33 malignant lymph nodes). The results showed a sensitivity of 0.80, a specificity of 0.67 and an accuracy of 0.74, but the class imbalance remains a significant challenge. This study highlights the potential of DW-MRI in oncology, while emphasizing the need for larger and more balanced datasets and more accurate ground truth images.
The segmentation and classification of lymph nodes in DW-MRI images are crucial for the diagnosis and monitoring of lymphoma, a tumor with a growing global incidence. DW-MRI, a non-invasive technique, exploits the different diffusion of water molecules to distinguish malignant tissues, which are denser and more compact, from benign ones, characterized by greater molecular mobility. Automatic segmentation, performed using convolutional neural networks, reduces time and errors compared to manual methods. In this study, on 45 patients, a segmentation model based on the No New Net architecture was developed, with combined inputs (ADC and b1000), achieving a mean Dice score of 0.52, limited by the small number of patients, suboptimal resolution and partially inaccurate ground truth images. For the classification, a Random Forest was trained on morphological and statistical features selected through PCA on 18 patients (1562 benign and 33 malignant lymph nodes). The results showed a sensitivity of 0.80, a specificity of 0.67 and an accuracy of 0.74, but the class imbalance remains a significant challenge. This study highlights the potential of DW-MRI in oncology, while emphasizing the need for larger and more balanced datasets and more accurate ground truth images.
File
| Nome file | Dimensione |
|---|---|
The thesis is not available. |
|