Pulmonary masses in dogs are uncommon neoplasms, but they are of growing interest in clinical practice thanks to advances in imaging technologies. However, manual segmentation of pulmonary lesions on CT images is time-consuming and subject to a certain degree of inter-operator variability. To overcome these limitations, this study explored the use of artificial intelligence to automate the segmentation of pulmonary masses in dogs, with the goal of providing a practical support tool for clinical practice. A retrospective study was conducted, collecting 187 CT studies from two university veterinary hospitals and a teleradiology service, selected from an initial pool of 403 cases. Cases with images affected by significant artifacts or with diffuse consolidations were excluded. Each scan, containing at least one pulmonary mass >2 cm, was manually segmented using the “3D Slicer” software to generate ground truth masks. The dataset was used to train an AI model based on the nnU-Net v2 framework, employing five-fold cross-validation. The model’s performance was then tested on an independent set of 30 cases. The algorithm demonstrated high accuracy, achieving a mean Dice Similarity Coefficient (DSC) of 0.91 and an Average Symmetric Surface Distance (ASSD) of 1.88 mm. The algorithm worked best on homogeneous lung masses with clear boundaries and regular margins, while intralesional mineralization or pleural effusion often reduced accuracy and sometimes caused partial segmentation errors. On the other hand, the model reliably avoided mistaking lung consolidation for a mass. Another important factor was the multicenter dataset, which included CT scans from many different institutions using various scanners and protocols. Although this variability could sometimes lower accuracy, it also made the model more robust and better suited to everyday clinical practice.
Le masse polmonari nel cane sono neoplasie poco comuni, ma di crescente interesse nella pratica clinica grazie all’evoluzione di tecnologie di imaging avanzate. Tuttavia, la segmentazione manuale delle lesioni polmonari su immagini TC è un processo dispendioso in termini di tempo, oltre che soggetto a una certa variabilità tra operatori. Per superare questi limiti, questo studio ha esplorato l’utilizzo dell’intelligenza artificiale per automatizzare la segmentazione delle masse polmonari nel cane con l’idea di fornire uno strumento di supporto concreto alla pratica clinica. È stato condotto uno studio retrospettivo che ha raccolto 187 studi TC da due ospedali veterinari universitari e un servizio di teleradiologia, selezionati da un pool iniziale di 403 casi. Sono stati esclusi i casi con immagini interessate da artefatti marcati o con consolidamenti diffusi. Ogni scansione, contenente almeno una massa polmonare >2 cm, è stata segmentata manualmente con il software “3d Slicer” per generare maschere di verità di base. Il dataset è stato utilizzato per addestrare un modello di IA basato sul framework nnU-Net v2, sfruttando una validazione incrociata a 5 livelli. Le prestazioni del modello sono state poi testate su un set indipendente di 30 casi. L’algoritmo ha dimostrando di aver raggiunto un’elevata accuratezza con Dice Similarity Coefficient (DSC) medio di 0,91 e una Average Symmetric Surface Distance (ASSD) di 1,88 mm. Le migliori performance sono state osservate nelle masse polmonari omogenee, ben delimitate e con margini regolari, mentre la presenza di mineralizzazioni intralesionali o di versamento pleurico tendeva a ridurre la precisione, in alcuni casi portando a parziali errori di segmentazione. Al contrario, il modello si è dimostrato affidabile nel non confondere le aree di consolidamento polmonare con le masse. Un ulteriore aspetto rilevante è la natura multicentrica del dataset, costituito da esami TC provenienti da numerose strutture e acquisiti con apparecchiature e protocolli differenti. Questa ampia variabilità, pur rappresentando una possibile fonte di riduzione dell’accuratezza, ha contribuito a rendere il modello più robusto e meglio adattabile alle condizioni reali della pratica clinica.
Sviluppo di un algoritmo di intelligenza artificiale per la segmentazione automatica di lesioni polmonari da immagini TC
POLONI, GIULIA
2024/2025
Abstract
Pulmonary masses in dogs are uncommon neoplasms, but they are of growing interest in clinical practice thanks to advances in imaging technologies. However, manual segmentation of pulmonary lesions on CT images is time-consuming and subject to a certain degree of inter-operator variability. To overcome these limitations, this study explored the use of artificial intelligence to automate the segmentation of pulmonary masses in dogs, with the goal of providing a practical support tool for clinical practice. A retrospective study was conducted, collecting 187 CT studies from two university veterinary hospitals and a teleradiology service, selected from an initial pool of 403 cases. Cases with images affected by significant artifacts or with diffuse consolidations were excluded. Each scan, containing at least one pulmonary mass >2 cm, was manually segmented using the “3D Slicer” software to generate ground truth masks. The dataset was used to train an AI model based on the nnU-Net v2 framework, employing five-fold cross-validation. The model’s performance was then tested on an independent set of 30 cases. The algorithm demonstrated high accuracy, achieving a mean Dice Similarity Coefficient (DSC) of 0.91 and an Average Symmetric Surface Distance (ASSD) of 1.88 mm. The algorithm worked best on homogeneous lung masses with clear boundaries and regular margins, while intralesional mineralization or pleural effusion often reduced accuracy and sometimes caused partial segmentation errors. On the other hand, the model reliably avoided mistaking lung consolidation for a mass. Another important factor was the multicenter dataset, which included CT scans from many different institutions using various scanners and protocols. Although this variability could sometimes lower accuracy, it also made the model more robust and better suited to everyday clinical practice.| File | Dimensione | Formato | |
|---|---|---|---|
|
Giulia_Poloni.pdf
Accesso riservato
Dimensione
3.44 MB
Formato
Adobe PDF
|
3.44 MB | Adobe PDF |
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/20.500.12608/94562