This experimental thesis aims to develop and test an artificial intelligence algorithm for the automatic segmentation of hepatic tumor lesions in dogs, using computed tomography (CT) images. A total of 171 CT scans containing hepatic lesions were analyzed, obtained through a search in the database of the Veterinary Teaching Hospital of the University of Padua and other sources. Manual segmentation of the lesions was performed using the open-source software 3D Slicer, allowing for precise delineation of the lesion boundaries. The dataset was subsequently divided into training, validation, and test sets, and the algorithm was trained using a U-Net convolutional neural network (CNN) model. The results obtained demonstrate a Dice Score of 0.85, indicating good performance in the automatic segmentation of tumor lesions. A thorough analysis revealed that the algorithm is particularly effective in recognizing large hepatic masses, with a Dice Score of 0.95, while it showed lower performance for smaller lesions, with a value of 0.4. These discrepancies may be attributed to factors such as the intrinsic complexity of identifying smaller lesions and the characteristics of the dataset used. Despite current limitations, such as the lack of a size-based differentiation of lesions, the obtained results highlight the potential of the algorithm for future clinical applications. This study aims to provide clinical support to veterinary practitioners, emphasizing that artificial intelligence should assist, not replace, human judgment in radiological practice.
Questa tesi sperimentale si propone di sviluppare e testare un algoritmo di intelligenza artificiale per la segmentazione automatica delle lesioni tumorali epatiche nel cane, utilizzando immagini di tomografia computerizzata (TC). Sono state analizzate un totale di 171 scansioni TC contenenti lesioni epatiche, ottenute attraverso una ricerca nel database dell’Ospedale Veterinario Didattico dell'Università di Padova e altre fonti. La segmentazione manuale delle lesioni è stata effettuata utilizzando il software open source 3D Slicer, permettendo di definire con precisione i contorni delle lesioni. Il dataset è stato poi suddiviso in set di addestramento, validazione e test, e l'algoritmo è stato addestrato utilizzando un modello di rete neurale convoluzionale (CNN) U-Net. I risultati ottenuti mostrano un Dice Score di 0,85, indicando buone performance nella segmentazione automatica delle lesioni tumorali. L'analisi approfondita ha rivelato che l'algoritmo è particolarmente efficace nel riconoscere masse epatiche di grandi dimensioni, con un Dice Score di 0,95, mentre ha mostrato prestazioni inferiori per le lesioni di dimensioni minori, con un valore di 0,4. Queste discrepanze potrebbero essere attribuite a fattori come la complessità intrinseca nell’identificazione di lesioni più piccole e le caratteristiche del dataset utilizzato. Nonostante i limiti attuali, come la mancanza di una suddivisione per dimensione delle lesioni, i risultati ottenuti evidenziano il potenziale dell'algoritmo per applicazioni cliniche future. Questo studio si propone di offrire un supporto clinico ai medici veterinari, sottolineando che l'intelligenza artificiale deve assistere, e non sostituire, il giudizio umano nella pratica radiologica.
Approccio AI-based per la segmentazione di tumori del fegato nei cani tramite tomografia computerizzata
QUARESIMA, EMMA
2023/2024
Abstract
This experimental thesis aims to develop and test an artificial intelligence algorithm for the automatic segmentation of hepatic tumor lesions in dogs, using computed tomography (CT) images. A total of 171 CT scans containing hepatic lesions were analyzed, obtained through a search in the database of the Veterinary Teaching Hospital of the University of Padua and other sources. Manual segmentation of the lesions was performed using the open-source software 3D Slicer, allowing for precise delineation of the lesion boundaries. The dataset was subsequently divided into training, validation, and test sets, and the algorithm was trained using a U-Net convolutional neural network (CNN) model. The results obtained demonstrate a Dice Score of 0.85, indicating good performance in the automatic segmentation of tumor lesions. A thorough analysis revealed that the algorithm is particularly effective in recognizing large hepatic masses, with a Dice Score of 0.95, while it showed lower performance for smaller lesions, with a value of 0.4. These discrepancies may be attributed to factors such as the intrinsic complexity of identifying smaller lesions and the characteristics of the dataset used. Despite current limitations, such as the lack of a size-based differentiation of lesions, the obtained results highlight the potential of the algorithm for future clinical applications. This study aims to provide clinical support to veterinary practitioners, emphasizing that artificial intelligence should assist, not replace, human judgment in radiological practice.File | Dimensione | Formato | |
---|---|---|---|
Quaresima_Emma.pdf
accesso aperto
Dimensione
1.49 MB
Formato
Adobe PDF
|
1.49 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/74349