Radiology is a frontline diagnostic tool widely used in small animal clinics, in particular for investigating the thorax. The evaluation of a radiographic image can be subjective and complex, especially when identifying suspicious masses, as their presentation can vary significantly. The radiographic characteristics of the main lesions that are or appear as masses, belonging to different thoracic regions, are discussed. With the introduction of artificial intelligence in veterinary medicine, various convolutional neural networks (CNNs) are being tested and utilized to analyze radiographic images. These networks are able to distinguish between pathological and normal radiographs, or recognize one or more present lesions. The goal is to standardize the diagnostic process, thereby supporting veterinarians in their daily practice. Studies conducted so far indicate that these algorithms generally show low accuracy in recognizing pulmonary masses in dog and cat radiographs. The aim is to compare these results with those obtained in the detection of thoracic masses in humans. The objective of this study is to understand how three variables - number, location, size – related to the masses, manually identified in canine latero thoracic radiographs, influence the performance of an artificial intelligence algorithm, specifically the V.E.R.A. software (Virtual Veterinary Radiolog Assistant). This software was developed by the Diagnostic Imaging team at the Veterinary Teaching Hospital of the University of Padua (O.V.U.D.) and is designed to detect the most common thoracic lesions in dogs. The algorithm performs better on images that present multiple masses. It detects the masses that are large on average 3.19 times the length of the T4 vertebrae, but it was not possible to find a dimensional cut off between the masses detected and not, because in some cases nodules have been correctly recognized, in others, major masses have not been detected. The size of the database did not allow to generalize the results obtained for the "position" parameter. With an overall error rate of 27.5% for this lesion and aware of the variability in mass presentation influencing the software performance, the model requires further implementation to be used in clinical practice with the aim of obtaining a second opinion, as the practitioner achieves better results.
La radiologia è un mezzo diagnostico di prima linea ampiamente impiegato nella clinica dei piccoli animali per investigarne il torace. La valutazione di un’immagine radiografica è soggettiva e spesso complessa in particolare quando si rilevano neoformazioni sospette per la loro variabilità di presentazione. Con la comparsa dell’intelligenza artificiale nell’ambito della medicina veterinaria diverse reti neurali convoluzionali (CNN) vengono testate e utilizzate per analizzare immagini radiografiche: distinguere una radiografia con o senza patologie oppure riconoscere una o più lesioni presenti. Lo scopo è quello di standardizzare il processo di rilevazione per affiancare il veterinario nella pratica quotidiana. Dagli studi fin’ora condotti questi algoritmi presentano una ridotta accuratezza nel riconoscimento delle masse polmonari nelle radiografie di cane e gatto. Si vuole confrontare questi risultati con quelli ottenuti nella valutazione delle radiografie toraciche dell’uomo. L’obiettivo dello studio è capire quali sono le variabili che incidono su queste scarse performance, nello specifico relative al software V.E.R.A. (Virtual Veterinary Radiology Assistant) sviluppato dall’equipe di Diagnostica per immagini dell’Ospedale didattico veterinario dell’Università degli Studi di Padova e capace di rilevare le più comuni lesioni toraciche del cane. Sono state valutate radiografie toraciche latero laterali di cane che presentano masse diverse per numero, posizione e dimensione. I primi due parametri risultano influenzare l’abilità di riconoscere la presenza di neoplasie polmonari.
Valutazione delle variabili che influenzano la capacità di un algoritmo di intelligenza artificiale di rilevare masse polmonari in radiografie toraciche latero laterali di cane
GIORDANO, MARTINA
2023/2024
Abstract
Radiology is a frontline diagnostic tool widely used in small animal clinics, in particular for investigating the thorax. The evaluation of a radiographic image can be subjective and complex, especially when identifying suspicious masses, as their presentation can vary significantly. The radiographic characteristics of the main lesions that are or appear as masses, belonging to different thoracic regions, are discussed. With the introduction of artificial intelligence in veterinary medicine, various convolutional neural networks (CNNs) are being tested and utilized to analyze radiographic images. These networks are able to distinguish between pathological and normal radiographs, or recognize one or more present lesions. The goal is to standardize the diagnostic process, thereby supporting veterinarians in their daily practice. Studies conducted so far indicate that these algorithms generally show low accuracy in recognizing pulmonary masses in dog and cat radiographs. The aim is to compare these results with those obtained in the detection of thoracic masses in humans. The objective of this study is to understand how three variables - number, location, size – related to the masses, manually identified in canine latero thoracic radiographs, influence the performance of an artificial intelligence algorithm, specifically the V.E.R.A. software (Virtual Veterinary Radiolog Assistant). This software was developed by the Diagnostic Imaging team at the Veterinary Teaching Hospital of the University of Padua (O.V.U.D.) and is designed to detect the most common thoracic lesions in dogs. The algorithm performs better on images that present multiple masses. It detects the masses that are large on average 3.19 times the length of the T4 vertebrae, but it was not possible to find a dimensional cut off between the masses detected and not, because in some cases nodules have been correctly recognized, in others, major masses have not been detected. The size of the database did not allow to generalize the results obtained for the "position" parameter. With an overall error rate of 27.5% for this lesion and aware of the variability in mass presentation influencing the software performance, the model requires further implementation to be used in clinical practice with the aim of obtaining a second opinion, as the practitioner achieves better results.File | Dimensione | Formato | |
---|---|---|---|
Giordano_Martina.pdf
accesso aperto
Dimensione
3.77 MB
Formato
Adobe PDF
|
3.77 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/20.500.12608/74344