Pathology localization on chest radiographs with limited supervision via semi-supervised multiple instance learning
Radiographs are the primary examination for diagnosing chest conditions, and yet they are frequently misread/misdiagnosed due to human-observer confusion. In clinical practice, there is an increase of deep learning approaches to support radiologists on the decision-making process to improve diagnost...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2021 |
| País: | Brasil |
| Institución: | Pontifícia Universidade Católica do Rio Grande do Sul (PUCRS) |
| Repositorio: | Biblioteca Digital de Teses e Dissertações da PUC_RS |
| Idioma: | inglés |
| OAI Identifier: | oai:tede2.pucrs.br:tede/9790 |
| Acceso en línea: | http://tede2.pucrs.br/tede2/handle/tede/9790 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep Learning Medical Imaging Semi-Supervised Learning Multiple Instance Learning Aprendizado Profundo Imagens Médicas Aprendizado Semi-Supervisionado Aprendizado de Múltiplas Instâncias CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO |
| Sumario: | Radiographs are the primary examination for diagnosing chest conditions, and yet they are frequently misread/misdiagnosed due to human-observer confusion. In clinical practice, there is an increase of deep learning approaches to support radiologists on the decision-making process to improve diagnostic accuracy. To properly support radiologists, it is insufficient for the system to simply output a diagnosis label. Ideally, the model should provide more information to support the classification result, such as the spatial localization of the finding. To properly train deep learning models, we usually need lots of annotated data. There is a vast amount of publicly-available chest radiographs labeled according to their radiological findings (labels for classification), but very few contain a location annotation. Our goal is to extend the use of unlabeled data to improve pathology localization in chest radiographs in a scenario with limited labeled data. We identify state-of-the-art semi-supervised methods and evaluated their performance on a classification scenario. Next, we extend the best method, Mean Teacher, to perform localization within a multiple instance learning framework, introducing our method C-MIL. Multiple instance learning is a paradigm with two types of labels: a general label that is known, and a more specific and unknown label but related to the one known, in our case, pathology presence and its localization. Our results show improvements of applying consistency regularization over a multiple instance localization framework and demonstrate that semi-supervised learning methods are promising to advance the state-of-the-art performance of pathology localization methods. |
|---|