Tracheal collapse is one of the main problems that plague dogs, especially those of smaller size and older age, causing symptoms that make life arduous for the animal and, in the worst cases, can lead to death. Therefore, the need for a quick and practical diagnosis of this problem in the animal as soon as it arrives at the clinic becomes fundamental. However, current diagnostic methods take a critical time by needing two different radiographs in the process, besides increasing the stress of the patient due to the need for external pressure on the dog's cervical trachea region during one of these radiographs. With that in mind, obtaining a tool for diagnosing cervical tracheal collapse in dogs with only basic patient data and only one radiographic exposure (the one without the need for compression of the animal's cervical trachea) is something that could contribute to a protocol that would allow a faster and more practical diagnosis of the disease, in addition to allowing the institution of early therapy, favoring a good prognosis for the patient. With that in mind, and the fact that Convolutional Neural Networks (CNNs) have shown results in the past in classification cases, as well as attention models (Transformers), especially in tests performed on humans, the aim of this project is to investigate the use of CNNs and Transformers to extract knowledge from radiographs, combined with a traditional Machine Learning model to bring extra information to the system from basic animal data, aiming at the development of a tool to aid in the explainable diagnosis of cervical tracheal collapse in dogs. Finally, it is expected that this tool can help the veterinarian by optimizing the diagnosis of cervical tracheal collapse in dogs, with a lower dose of radiation for both, professional and animal, in addition to reducing diagnostic costs, all direct results of the possibility of reducing an radiographic exposure.
News published in Agência FAPESP Newsletter about the scholarship: