Loading...
Publication
Segmentation of wounds and pressure ulcers
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 11.45 MB | Adobe PDF |
Authors
Abstract(s)
Os profissionais de saúde enfrentam desafios ao avaliar e segmentar feridas em imagens
médicas, uma tarefa essencial para controlar a evolução da cicatrização e planear
tratamentos adequados. O presente trabalho propõe a aplicação de modelos de deep
learning para a segmentação de feridas, explorando duas arquiteturas principais: YOLOv9
e Medical Transformer (MedT).Oobjetivo é identificar a melhor abordagem para
automatizar a segmentação de diferentes tipos de feridas, reduzindo a subjetividade e
o tempo associados aos métodos tradicionais. Para posteriormente obter as dimensões
das feridas.
Ao realizar este estudo, foi preparado um conjunto de dados que combina imagens
recolhidas de fontes online e de uma parceria com uma instituição de saúde. Essas
imagens foram cuidadosamente anotadas para treinar os modelos propostos. A YOLOv9
destacou-se pela eficiência na identificação de marcas de calibração, enquanto
o MedT demonstrou maior consistência na segmentação de feridas, aproveitando as
capacidades de atenção global dos transformers.
Os resultados obtidos mostram que, apesar dos desafios, como falsos positivos e limitações
do conjunto de dados, os modelos apresentam potencial significativo para
melhorar a avaliação clínica de feridas. Este projeto representa um avanço na aplicação
de inteligência artificial na área da saúde, contribuindo para o desenvolvimento de
sistemas automáticos de segmentação que podem servir como ferramentas de apoio à
decisão para profissionais de saúde.
Healthcare professionals face challenges when assessing and segmenting wounds in medical images, an essential task for monitoring the progress of healing and planning appropriate treatments. This paper proposes the application of deep learning models for wound segmentation, exploring two main architectures: YOLOv9 and Medical Transformer (MedT). The aim is to identify the best approach for automating the segmentation of different types of wounds, reducing the subjectivity and time associated with traditional methods. In order to subsequently obtain the dimensions of the wounds. To carry out this study, a datasetwas prepared combining images collected from online sources and from a partnership with a healthcare institution. These images were carefully annotated to train the proposed models. YOLOv9 stood out for its efficiency in identifying calibration marks, while MedT showed greater consistency in segmenting wounds, taking advantage of the transformers’ global attention capabilities. The results obtained show that, despite challenges such as false positives and limitations of the data set, the models have significant potential to improve the clinical assessment of wounds. This project represents a breakthrough in the application of artificial intelligence in healthcare, contributing to the development of automated segmentation systems that can serve as decision support tools for healthcare professionals.
Healthcare professionals face challenges when assessing and segmenting wounds in medical images, an essential task for monitoring the progress of healing and planning appropriate treatments. This paper proposes the application of deep learning models for wound segmentation, exploring two main architectures: YOLOv9 and Medical Transformer (MedT). The aim is to identify the best approach for automating the segmentation of different types of wounds, reducing the subjectivity and time associated with traditional methods. In order to subsequently obtain the dimensions of the wounds. To carry out this study, a datasetwas prepared combining images collected from online sources and from a partnership with a healthcare institution. These images were carefully annotated to train the proposed models. YOLOv9 stood out for its efficiency in identifying calibration marks, while MedT showed greater consistency in segmenting wounds, taking advantage of the transformers’ global attention capabilities. The results obtained show that, despite challenges such as false positives and limitations of the data set, the models have significant potential to improve the clinical assessment of wounds. This project represents a breakthrough in the application of artificial intelligence in healthcare, contributing to the development of automated segmentation systems that can serve as decision support tools for healthcare professionals.
Description
Keywords
Feridas Segmentação Deep learning Rede neuronais