A carregar...
Publicação
Aplicação da inteligência artificial em radiografias para o diagnóstico de derrame pleural em gatos
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 6.07 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A aplicação de Inteligência Artificial (IA) no diagnóstico por imagem é um tópico de interesse crescente na medicina veterinária. Machine Learning (ML) é um ramo da IA, cujas ferramentas permitem a deteção de padrões significativos a partir de dados, incluindo imagens, com capacidade superior à de um observador humano. A tecnologia de ML pode ser aplicada na radiologia para o diagnóstico de doenças que se traduzem em alterações radiográficas, como o derrame pleural. Este trabalho teve como objetivo desenvolver dois modelos de ML para classificação de imagens radiográficas, para o diagnóstico de derrame pleural em gatos, recorrendo a uma plataforma de Machine Learning automatizado, a Vertex AI: o modelo A, um modelo binário para identificação da presença ou ausência de derrame pleural (rótulos: “YesDerrame”, NoDerrame”, respetivamente), e o modelo B, um modelo de classificação multiclasse para quantificação do volume de derrame pleural (rótulos: “Ausente”, “NãoDrenável”, “Drenável”). As imagens foram recolhidas no Hospital Escolar Veterinário da Faculdade de Medicina Veterinária da Universidade de Lisboa, Portugal. Os casos positivos incluem a confirmação de derrame pleural com um segundo método de diagnóstico, e os casos negativos não apresentam sinais clínicos nem radiográficos de derrame pleural. Parte das imagens de gatos foi utilizada para desenvolver os modelos A e B (70% treino, 20% validação) e a restante para testar a sua performance (10%). As imagens de treino e validação foram submetidas a data augmentation (rotação e inclinação), obtendo-se 9072 novas imagens, e este conjunto foi utilizado para treinar uma segunda versão do modelo A. As imagens de cães foram utilizadas apenas para testar as versões do modelo A, para avaliar a capacidade de generalização do modelo. Foram selecionadas 318 imagens (280 de gatos e 38 de cães), para desenvolvimento e testagem dos modelos. As quatro tarefas de testagem do modelo A apresentaram bons parâmetros de performance (≥80%) e não se verificaram diferenças estatisticamente significativas entre elas. No modelo B, os parâmetros de performance no geral foram aceitáveis (> 75,0%), no entanto, a performance para o rótulo “NãoDrenável” (precisão (VPP) 54,4% e F1-score <70%) evidencia a ocorrência de underfitting no treino do modelo. Concluiu-se que o modelo A tem potencial para aplicação em contexto clínico, sendo necessária uma nova testagem com um maior número de imagens para validar com clareza a sua performance. O modelo B não é viável para a quantificação do volume de derrame pleural, devido a underfitting, mas pode ser útil para confirmar a indicação para toracocentese e para sinalizar imagens que devem ser interpretadas pelo médico veterinário
ABSTRACT - APPLICATION OF ARTIFICIAL INTELLIGENCE IN RADIOGRAPHIC IMAGES FOR THE DIAGNOSIS OF PLEURAL EFFUSION IN CATS - The application of artificial intelligence in diagnostic imaging is a topic of increasing interest in veterinary medicine. Machine Learning (ML), a field of artificial intelligence, enables the detection of significant patterns in data, including images, with greater capacity than a human observer. ML technology can be applied in radiology for the diagnosis of diseases that present radiographic signs such as pleural effusion. The goal of this paper was to develop two Machine Learning models for the classification of radiographic images, for the diagnosis of pleural effusion in cats, using an automated Machine Learning platform, Vertex AI: model A, a binary classification model for the presence or absence of pleural effusion (ground truth: “YesDerrame”, “NoDerrame”, respectively), and model B, a multiclass classification model for the quantification of the volume of pleural effusion (ground truth: “Ausente”, “NãoDrenável”, “Drenável”, meaning "Absent”, “Non-drainable” and “Drainable”, respectively). The radiographic images were collected from the Faculty of Veterinary Medicine’s Teaching Hospital, University of Lisbon, Portugal, with the positive cases including confirmation of pleural effusion through a second diagnostic method, while the negative cases did not present any clinical or radiographic signs of pleural effusion. A portion of the cat images was used for the evelopment of models A and B (70% training, 20% validation), and the remaining 10% for assessing the models’ performance. The training and validation images were submitted to data augmentation (rotation and shearing), resulting in 9072 new images, which were used to train a new version of model A. The images of dogs were only used for testing the versions of model A, to evaluate their generalization ability. A total of 318 images were selected (280 cats and 38 dogs), for the models’ development and testing. All four of model A’s testing tasks presented performance metrics equal to or higher than 80%, and there were no statistically significant differences between them. For model B, the performance metrics were acceptable overall (>75.0%); however, the performance for the “Non-drainable” label (precision 54.4% and F1-score <70%) provides evidence of underfitting in the training of the model. In conclusion, model A has potential for application in a clinical setting, but further testing with a larger dataset is necessary to improve confidence in its predictions. Model B, however, is not viable for quantifying pleural effusion volume due to underfitting but could be useful for confirming the indication for thoracocentesis in large volume pleural effusions, or for flagging images where an interpretation by the veterinarian is required
ABSTRACT - APPLICATION OF ARTIFICIAL INTELLIGENCE IN RADIOGRAPHIC IMAGES FOR THE DIAGNOSIS OF PLEURAL EFFUSION IN CATS - The application of artificial intelligence in diagnostic imaging is a topic of increasing interest in veterinary medicine. Machine Learning (ML), a field of artificial intelligence, enables the detection of significant patterns in data, including images, with greater capacity than a human observer. ML technology can be applied in radiology for the diagnosis of diseases that present radiographic signs such as pleural effusion. The goal of this paper was to develop two Machine Learning models for the classification of radiographic images, for the diagnosis of pleural effusion in cats, using an automated Machine Learning platform, Vertex AI: model A, a binary classification model for the presence or absence of pleural effusion (ground truth: “YesDerrame”, “NoDerrame”, respectively), and model B, a multiclass classification model for the quantification of the volume of pleural effusion (ground truth: “Ausente”, “NãoDrenável”, “Drenável”, meaning "Absent”, “Non-drainable” and “Drainable”, respectively). The radiographic images were collected from the Faculty of Veterinary Medicine’s Teaching Hospital, University of Lisbon, Portugal, with the positive cases including confirmation of pleural effusion through a second diagnostic method, while the negative cases did not present any clinical or radiographic signs of pleural effusion. A portion of the cat images was used for the evelopment of models A and B (70% training, 20% validation), and the remaining 10% for assessing the models’ performance. The training and validation images were submitted to data augmentation (rotation and shearing), resulting in 9072 new images, which were used to train a new version of model A. The images of dogs were only used for testing the versions of model A, to evaluate their generalization ability. A total of 318 images were selected (280 cats and 38 dogs), for the models’ development and testing. All four of model A’s testing tasks presented performance metrics equal to or higher than 80%, and there were no statistically significant differences between them. For model B, the performance metrics were acceptable overall (>75.0%); however, the performance for the “Non-drainable” label (precision 54.4% and F1-score <70%) provides evidence of underfitting in the training of the model. In conclusion, model A has potential for application in a clinical setting, but further testing with a larger dataset is necessary to improve confidence in its predictions. Model B, however, is not viable for quantifying pleural effusion volume due to underfitting but could be useful for confirming the indication for thoracocentesis in large volume pleural effusions, or for flagging images where an interpretation by the veterinarian is required
Descrição
Dissertação de Mestrado Integrado em Medicina Veterinária, área científica de Clínica
Palavras-chave
Derrame pleural Gatos Inteligência artificial Machine learning Radiografia Artificial intelligence Cats Machine learning Pleural effusion Radiography
Contexto Educativo
Citação
Guerra MM. 2025. Aplicação da inteligência artificial em radiografias para o diagnóstico de derrame pleural em gatos [dissertação de mestrado]. Lisboa: FMV-Universidade de Lisboa
Editora
Universidade de Lisboa, Faculdade de Medicina Veterinária