Publicação
VCE dataset generation: active learning solutions for binary classification in informative vs uninformative frames
| Resumo: | Video Capsule Endoscopy is a non-invasive image technique that allows the observation of the small bowel. However, it requires review and Annotation of up to 8 to 10 hours of videos that need to be reviewed by a medical expert, which is very time-consuming. State-of-the-art Machine Learning methods now have the power to assist experts by automatically classifying findings in the video frames, but big Video Capsule Endoscopy annotated datasets are needed, which requires an unaffordable effort. Active Learning methodologies can be used to optimize dataset annotation through the intelligent identification of the samples to be annotated in big non-annotated datasets that most contribute to model learning. In this dissertation, a study of Active Learning to create VCE datasets, in order to solve a binary problem related to the classification between informative and uninformative frames, was made. We explored some Active Learning techniques, such as Least Confidence Sampling and Margin Sampling, to conclude about the annotation effort and the capability to rapidly create representative datasets. It was verified that Least Confidence Sampling was the more appropriate technique for our data, given the accuracy when dividing unseen video frames into informative and uninformative; and that Active Learning has the potential to expand the existing datasets using less data and human effort. |
|---|---|
| Autores principais: | Nunes, Beatriz Gramata |
| Assunto: | VCE Active learning Dataset creation Informative images |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Aveiro |
| Idioma: | inglês |
| Origem: | RIA - Repositório Institucional da Universidade de Aveiro |
| Resumo: | Video Capsule Endoscopy is a non-invasive image technique that allows the observation of the small bowel. However, it requires review and Annotation of up to 8 to 10 hours of videos that need to be reviewed by a medical expert, which is very time-consuming. State-of-the-art Machine Learning methods now have the power to assist experts by automatically classifying findings in the video frames, but big Video Capsule Endoscopy annotated datasets are needed, which requires an unaffordable effort. Active Learning methodologies can be used to optimize dataset annotation through the intelligent identification of the samples to be annotated in big non-annotated datasets that most contribute to model learning. In this dissertation, a study of Active Learning to create VCE datasets, in order to solve a binary problem related to the classification between informative and uninformative frames, was made. We explored some Active Learning techniques, such as Least Confidence Sampling and Margin Sampling, to conclude about the annotation effort and the capability to rapidly create representative datasets. It was verified that Least Confidence Sampling was the more appropriate technique for our data, given the accuracy when dividing unseen video frames into informative and uninformative; and that Active Learning has the potential to expand the existing datasets using less data and human effort. |
|---|