In recent decades, the increasing global spread of brachycephalic breeds has made it essential for veterinarians to gain a thorough knowledge of their morphological characteristics and the pathologies to which they are predisposed. The particular shape of these breeds is reflected in the radiographic images, which are difficult to interpret. To speed up the diagnostic process and limit operator-dependent errors, VERA (Virtual Veterinary Radiology Assistant) was developed, a platform that uses a pre-trained convolutional deep neural network (ResNet-50) to automatically detect lesions in canine chest radiographs. This study evaluated the effectiveness of the algorithm in classifying radiographs of brachycephalic dogs, analyzing 105 chest-lateral radiographs of dogs belonging to 14 brachycephalic breeds sampled from the University Veterinary Teaching Hospital of the University of Padua database. First, these images were classified manually by applying diagnostic labels, then, the same images were classified automatically by VERA. The performance of the algorithm was evaluated in sensitivity, specificity, positive likelihood ratio (PLR) and negative likelihood ratio (NLR) for each condition found. These statistical indices were compared with those calculated from a dataset of 776 chest lateral-lateral radiographs of different dog breeds, previously evaluated in the study "Automatic classification of canine thoracic radiographs using deep learning" (Banzato et al., 2021). The analysis of the results revealed that VERA performs moderately well in detecting lesions in chest radiographs of brachycephalic dogs, although it performs worse when compared to the results for the general population dataset. This suggests a difficulty of the deep neural network in detecting lesions in X-rays of these specific breeds. Therefore, VERA’s reliability is not sufficient to justify its use as the only source of diagnosis, which always requires the validation of the veterinarian, who must be able to interpret the radiographic peculiarities of brachycephalic breeds. Nevertheless, VERA has shown excellent ability to correctly rule out lesions, confirming its usefulness as a clinical support tool.
Negli ultimi decenni la crescente diffusione delle razze brachicefale a livello globale ha reso essenziale per i medici veterinari acquisire una conoscenza approfondita delle loro caratteristiche morfologiche e delle patologie a cui sono predisposti. La particolare conformazione di queste razze si riflette nelle immagini radiografiche, che risultano essere complicate da interpretare. Per velocizzare il processo diagnostico e limitare gli errori operatore-dipendenti, è stata sviluppata VERA (Virtual Veterinary Radiology Assistant), una piattaforma che utilizza una rete neurale convoluzionale pre-addestrata (ResNet-50) per rilevare automaticamente la presenza di lesioni nelle radiografie toraciche canine. Questo studio ha valutato l’efficacia dell’algoritmo nella classificazione delle radiografie di cani brachicefali, analizzando 105 radiografie toraciche latero-laterali di cani appartenenti a 14 razze brachicefale, campionate da un database di immagini provenienti dall’Ospedale Veterinario Universitario Didattico dell’Università di Padova. Alla classificazione manuale di queste radiografie, tramite l’applicazione di etichette diagnostiche, è seguita quella automatica di VERA. La performance dell’algoritmo è stata valutata in termini di sensibilità, specificità, positive likelihood ratio (PLR) e negative likelihood ratio (NLR) per ciascuna lesione riscontrata. Questi indici statistici sono stati confrontati con quelli calcolati per un dataset composto di 776 radiografie toraciche latero-laterali di diverse razze canine, precedentemente valutato nello studio "Automatic classification of canine thoracic radiographs using deep learning" (Banzato et al., 2021). L'analisi dei risultati ha rivelato che VERA offre prestazioni moderatamente buone nell’individuazione delle lesioni nelle radiografie toraciche di cani brachicefali, ma notevolmente inferiori rispetto a quelle ottenute per la popolazione generale. Questo suggerisce una difficoltà della rete neurale nel rilevare lesioni nelle radiografie di queste razze specifiche. L’affidabilità di VERA, quindi, non è sufficiente a giustificarne l'uso come unica fonte di diagnosi, la quale richiede sempre la verifica del medico veterinario, che deve conoscere e saper interpretare le peculiarità radiografiche delle razze brachicefale. Tuttavia, VERA ha dimostrato un'ottima capacità di escludere correttamente la presenza di lesioni, confermandone l’utilità come strumento di supporto clinico.
Valutazione della performance di un algoritmo di intelligenza artificiale nella rilevazione di lesioni toraciche in radiografie latero-laterali di un campione di cani brachicefali rispetto alla popolazione generale
LANZARINI, ANNA
2023/2024
Abstract
In recent decades, the increasing global spread of brachycephalic breeds has made it essential for veterinarians to gain a thorough knowledge of their morphological characteristics and the pathologies to which they are predisposed. The particular shape of these breeds is reflected in the radiographic images, which are difficult to interpret. To speed up the diagnostic process and limit operator-dependent errors, VERA (Virtual Veterinary Radiology Assistant) was developed, a platform that uses a pre-trained convolutional deep neural network (ResNet-50) to automatically detect lesions in canine chest radiographs. This study evaluated the effectiveness of the algorithm in classifying radiographs of brachycephalic dogs, analyzing 105 chest-lateral radiographs of dogs belonging to 14 brachycephalic breeds sampled from the University Veterinary Teaching Hospital of the University of Padua database. First, these images were classified manually by applying diagnostic labels, then, the same images were classified automatically by VERA. The performance of the algorithm was evaluated in sensitivity, specificity, positive likelihood ratio (PLR) and negative likelihood ratio (NLR) for each condition found. These statistical indices were compared with those calculated from a dataset of 776 chest lateral-lateral radiographs of different dog breeds, previously evaluated in the study "Automatic classification of canine thoracic radiographs using deep learning" (Banzato et al., 2021). The analysis of the results revealed that VERA performs moderately well in detecting lesions in chest radiographs of brachycephalic dogs, although it performs worse when compared to the results for the general population dataset. This suggests a difficulty of the deep neural network in detecting lesions in X-rays of these specific breeds. Therefore, VERA’s reliability is not sufficient to justify its use as the only source of diagnosis, which always requires the validation of the veterinarian, who must be able to interpret the radiographic peculiarities of brachycephalic breeds. Nevertheless, VERA has shown excellent ability to correctly rule out lesions, confirming its usefulness as a clinical support tool.File | Dimensione | Formato | |
---|---|---|---|
Lanzarini_Anna.pdf
accesso riservato
Dimensione
3.59 MB
Formato
Adobe PDF
|
3.59 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/20.500.12608/70935